Programming in Python 3
Programming in Python 3
A Complete Introduction to the Python Language
Second Edition
Mark Summerfield
Upper Saddle River, NJ · Boston · Indianapolis · San Francisco
New York · Toronto · Montreal · London · Munich · Paris · Madrid
Capetown · Sydney · Tokyo · Singapore · Mexico City
Many of the designations used by manufacturers and sellers to distinguish their products are claimed
as trademarks. Where those designations appear in this book, and the publisher was aware of a
trademark claim, the designations have been printed with initial capital letters or in all capitals.
The author and publisher have taken care in the preparation of this book, but make no expressed or
implied warranty of any kind and assume no responsibility for errors or omissions. No liability is
assumed for incidental or consequential damages in connection with or arising out of the use of the
information or programs contained herein.
The publisher offers excellent discounts on this book when ordered in quantity for bulk purchases
or special sales, which may include electronic versions and/or custom covers and content particular
to your business, training goals, marketing focus, and branding interests. For more information,
please contact:
U.S. Corporate and Government Sales
(800) 382-3419
[email protected]
For sales outside the United States, please contact:
International Sales
[email protected]
Visit us on the Web:
Library of Congress Cataloging-in-Publication Data
Summerfield, Mark.
Programming in Python 3 : a complete introduction to the Python language / Mark
Summerfield.—2nd ed.
p. cm.
Includes bibliographical references and index.
ISBN 978-0-321-68056-3 (pbk. : alk. paper)
1. Python (Computer program language) 2. Object-oriented programming (Computer science)
I. Title.
QA76.73.P98S86 2010
Copyright © 2010 Pearson Education, Inc.
All rights reserved. Printed in the United States of America. This publication is protected by
copyright, and permission must be obtained from the publisher prior to any prohibited reproduction,
storage in a retrieval system, or transmission in any form or by any means, electronic, mechanical,
photocopying, recording, or likewise. For information regarding permissions, write to:
Pearson Education, Inc.
Rights and Contracts Department
501 Boylston Street, Suite 900
Boston, MA 02116
Fax: (617) 671-3447
ISBN-13: 978-0-321-68056-3
Text printed in the United States on recycled paper at RR Donnelley in Crawfordsville, Indiana.
First printing, November 2009
Listof Tables
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter1. RapidIntroductiontoProceduralProgramming
Chapter2. DataTypes
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Chapter3. CollectionDataTypes
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
. . . . . . . . . . . . . . . . . . . 159
Chapter4. ControlStructuresandFunctions
Chapter5. Modules
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Chapter6. Object-OrientedProgramming
Chapter7. FileHandling
. . . . . . . . . . . . . . . . . . . . . . 233
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Chapter8. AdvancedProgrammingTechniques
Chapter9. Debugging,Testing,andProfiling
Chapter10. ProcessesandThreading
Chapter11. Networking
. . . . . . . . . . . . . . . . 339
. . . . . . . . . . . . . . . . . . . 413
. . . . . . . . . . . . . . . . . . . . . . . . . . . 439
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Chapter12. DatabaseProgramming
Chapter13. RegularExpressions
Chapter14. IntroductiontoParsing
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Chapter15. IntroductiontoGUIProgramming
. . . . . . . . . . . . . . . . . 569
Epilogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
SelectedBibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 1. Rapid Introduction to Procedural Programming . . .
Creating and Running Python Programs . . . . . . . . . . . . . . . . . . . . . . . .
Python’s “Beautiful Heart” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #1: Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #2: Object References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #3: Collection Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #4: Logical Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #5: Control Flow Statements . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #6: Arithmetic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #7: Input/Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Piece #8: Creating and Calling Functions . . . . . . . . . . . . . . . . . . . .
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 2. Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Identifiers and Keywords . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Integral Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Booleans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Floating-Point Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Floating-Point Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Complex Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Decimal Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparing Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Slicing and Striding Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
String Operators and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
String Formatting with the str.format() Method . . . . . . . . . . . . . .
Character Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 3. Collection Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sequence Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Tuples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Named Tuples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Set Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Frozen Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mapping Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dictionaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Default Dictionaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ordered Dictionaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Iterating and Copying Collections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Iterators and Iterable Operations and Functions . . . . . . . . . . . . .
Copying Collections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 4. Control Structures and Functions . . . . . . . . . . . . . . . . . . .
Control Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conditional Branching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Looping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Catching and Raising Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . .
Custom Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Custom Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Names and Docstrings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Argument and Parameter Unpacking . . . . . . . . . . . . . . . . . . . . . . .
Accessing Variables in the Global Scope . . . . . . . . . . . . . . . . . . . . .
Lambda Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Assertions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 5. Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Modules and Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Custom Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Overview of Python’s Standard Library . . . . . . . . . . . . . . . . . . . . . . . . . .
String Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Command-Line Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mathematics and Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Times and Dates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Algorithms and Collection Data Types . . . . . . . . . . . . . . . . . . . . . . .
File Formats, Encodings, and Data Persistence . . . . . . . . . . . . . . .
File, Directory, and Process Handling . . . . . . . . . . . . . . . . . . . . . . . .
Networking and Internet Programming . . . . . . . . . . . . . . . . . . . . .
XML . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Other Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 6. Object-Oriented Programming . . . . . . . . . . . . . . . . . . . . . .
The Object-Oriented Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Object-Oriented Concepts and Terminology . . . . . . . . . . . . . . . . . .
Custom Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Attributes and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Inheritance and Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using Properties to Control Attribute Access . . . . . . . . . . . . . . . .
Creating Complete Fully Integrated Data Types . . . . . . . . . . . . .
Custom Collection Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Classes That Aggregate Collections . . . . . . . . . . . . . . . .
Creating Collection Classes Using Aggregation . . . . . . . . . . . . . .
Creating Collection Classes Using Inheritance . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 7. File Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Writing and Reading Binary Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pickles with Optional Compression . . . . . . . . . . . . . . . . . . . . . . . . . .
Raw Binary Data with Optional Compression . . . . . . . . . . . . . . .
Writing and Parsing Text Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Writing Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing Text Using Regular Expressions . . . . . . . . . . . . . . . . . . . .
Writing and Parsing XML Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Element Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
DOM (Document Object Model) . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Manually Writing XML . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing XML with SAX (Simple API for XML) . . . . . . . . . . . . . . .
Random Access Binary Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A Generic BinaryRecordFile Class . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: The BikeStock Module’s Classes . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 8. Advanced Programming Techniques . . . . . . . . . . . . . . . .
Further Procedural Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Branching Using Dictionaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Generator Expressions and Functions . . . . . . . . . . . . . . . . . . . . . . .
Dynamic Code Execution and Dynamic Imports . . . . . . . . . . . . . .
Local and Recursive Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Function and Method Decorators . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Function Annotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Further Object-Oriented Programming . . . . . . . . . . . . . . . . . . . . . . . . . .
Controlling Attribute Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Functors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Context Managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Class Decorators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Abstract Base Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Multiple Inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Metaclasses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Functional-Style Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Partial Function Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Coroutines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 9. Debugging, Testing, and Profiling . . . . . . . . . . . . . . . . . . .
Debugging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dealing with Syntax Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dealing with Runtime Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Scientific Debugging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Unit Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 10. Processes and Threading . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using the Multiprocessing Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using the Threading Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: A Threaded Find Word Program . . . . . . . . . . . . . . . . . . .
Example: A Threaded Find Duplicate Files Program . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 11. Networking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating a TCP Client . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating a TCP Server . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 12. Database Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
DBM Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
SQL Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 13. Regular Expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Python’s Regular Expression Language . . . . . . . . . . . . . . . . . . . . . . . . . .
Characters and Character Classes . . . . . . . . . . . . . . . . . . . . . . . . . .
Quantifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Grouping and Capturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Assertions and Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The Regular Expression Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 14. Introduction to Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
BNF Syntax and Parsing Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . .
Writing Handcrafted Parsers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Simple Key–Value Data Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Playlist Data Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing the Blocks Domain-Specific Language . . . . . . . . . . . . . . .
Pythonic Parsing with PyParsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A Quick Introduction to PyParsing . . . . . . . . . . . . . . . . . . . . . . . . . .
Simple Key–Value Data Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Playlist Data Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing the Blocks Domain-Specific Language . . . . . . . . . . . . . . .
Parsing First-Order Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Lex/Yacc-Style Parsing with PLY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Simple Key–Value Data Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Playlist Data Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing the Blocks Domain-Specific Language . . . . . . . . . . . . . . .
Parsing First-Order Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 15. Introduction to GUI Programming . . . . . . . . . . . . . . . . .
Dialog-Style Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Main-Window-Style Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating a Main Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating a Custom Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Epilogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Selected Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
List of Tables
Python’s Keywords . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Numeric Operators and Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Integer Conversion Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Integer Bitwise Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
The Math Module’s Functions and Constants #1 . . . . . . . . . . . . . . 60
The Math Module’s Functions and Constants #2 . . . . . . . . . . . . . . 61
Python’s String Escapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
String Methods #1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
String Methods #2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
String Methods #3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
List Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Set Methods and Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Dictionary Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Common Iterable Operators and Functions . . . . . . . . . . . . . . . . . . . 140
Comparison Special Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
Fundamental Special Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
Numeric and Bitwise Special Methods . . . . . . . . . . . . . . . . . . . . . . . 253
Collection Special Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Bytes and Bytearray Methods #1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
Bytes and Bytearray Methods #2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
Bytes and Bytearray Methods #3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
File Object Attributes and Methods #1 . . . . . . . . . . . . . . . . . . . . . . . 325
File Object Attributes and Methods #2 . . . . . . . . . . . . . . . . . . . . . . . 326
Dynamic Programming and Introspection Functions . . . . . . . . . . 349
Attribute Access Special Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
The Numbers Module’s Abstract Base Classes . . . . . . . . . . . . . . . . 381
The Collections Module’s Main Abstract Base Classes . . . . . . . . . 383
DB-API 2.0 Connection Object Methods . . . . . . . . . . . . . . . . . . . . . . 481
DB-API 2.0 Cursor Object Attributes and Methods . . . . . . . . . . . 482
Character Class Shorthands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
Regular Expression Quantifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
Regular Expression Assertions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
The Regular Expression Module’s Functions . . . . . . . . . . . . . . . . . 502
The Regular Expression Module’s Flags . . . . . . . . . . . . . . . . . . . . . . 502
Regular Expression Object Methods . . . . . . . . . . . . . . . . . . . . . . . . . 503
Match Object Attributes and Methods . . . . . . . . . . . . . . . . . . . . . . . . 507
Python is probably the easiest-to-learn and nicest-to-use programming language in widespread use. Python code is clear to read and write, and it is concise without being cryptic. Python is a very expressive language, which means
that we can usually write far fewer lines of Python code than would be required
for an equivalent application written in, say, C++ or Java.
Python is a cross-platform language: In general, the same Python program can
be run on Windows and Unix-like systems such as Linux, BSD, and Mac OS X,
simply by copying the file or files that make up the program to the target
machine, with no “building” or compiling necessary. It is possible to create
Python programs that use platform-specific functionality, but this is rarely
necessary since almost all of Python’s standard library and most third-party
libraries are fully and transparently cross-platform.
One of Python’s great strengths is that it comes with a very complete standard
library—this allows us to do such things as download a file from the Internet,
unpack a compressed archive file, or create a web server, all with just one or a
few lines of code. And in addition to the standard library, thousands of thirdparty libraries are available, some providing more powerful and sophisticated facilities than the standard library—for example, the Twisted networking
library and the NumPy numeric library—while others provide functionality
that is too specialized to be included in the standard library—for example, the
SimPy simulation package. Most of the third-party libraries are available from
the Python Package Index,
Python can be used to program in procedural, object-oriented, and to a lesser
extent, in functional style, although at heart Python is an object-oriented
language. This book shows how to write both procedural and object-oriented
programs, and also teaches Python’s functional programming features.
The purpose of this book is to show you how to write Python programs in good
idiomatic Python 3 style, and to be a useful reference for the Python 3 language
after the initial reading. Although Python 3 is an evolutionary rather than revolutionary advance on Python 2, some older practices are no longer appropriate
or necessary in Python 3, and new practices have been introduced to take advantage of Python 3 features. Python 3 is a better language than Python 2—it
builds on the many years of experience with Python 2 and adds lots of new
features (and omits Python 2’s misfeatures), to make it even more of a pleasure
to use than Python 2, as well as more convenient, easier, and more consistent.
The book’s aim is to teach the Python language, and although many of the
standard Python libraries are used, not all of them are. This is not a problem,
because once you have read the book, you will have enough Python knowledge
to be able to make use of any of the standard libraries, or any third-party
Python library, and be able to create library modules of your own.
The book is designed to be useful to several different audiences, including selftaught and hobbyist programmers, students, scientists, engineers, and others
who need to program as part of their work, and of course, computing professionals and computer scientists. To be of use to such a wide range of people
without boring the knowledgeable or losing the less-experienced, the book assumes at least some programming experience (in any language). In particular, it assumes a basic knowledge of data types (such as numbers and strings),
collection data types (such as sets and lists), control structures (such as if and
while statements), and functions. In addition, some examples and exercises
assume a basic knowledge of HTML markup, and some of the more specialized
chapters at the end assume a basic knowledge of their subject area; for example, the databases chapter assumes a basic knowledge of SQL.
The book is structured in such a way as to make you as productive as possible
as quickly as possible. By the end of the first chapter you will be able to write
small but useful Python programs. Each successive chapter introduces new
topics, and often both broadens and deepens the coverage of topics introduced
in earlier chapters. This means that if you read the chapters in sequence,
you can stop at any point and you’ll be able to write complete programs with
what you have learned up to that point, and then, of course, resume reading
to learn more advanced and sophisticated techniques when you are ready. For
this reason, some topics are introduced in one chapter, and then are explored
further in one or more later chapters.
Two key problems arise when teaching a new programming language. The
first is that sometimes when it is necessary to teach one particular concept,
that concept depends on another concept, which in turn depends either directly
or indirectly on the first. The second is that, at the beginning, the reader may
know little or nothing of the language, so it is very difficult to present interesting or useful examples and exercises. In this book, we seek to solve both
of these problems, first by assuming some prior programming experience, and
second by presenting Python’s “beautiful heart” in Chapter 1—eight key pieces
of Python that are sufficient on their own to write decent programs. One consequence of this approach is that in the early chapters some of the examples
are a bit artificial in style, since they use only what has been taught up to the
point where they are presented; this effect diminishes chapter by chapter, until
by the end of Chapter 7, all the examples are written in completely natural and
idiomatic Python 3 style.
The book’s approach is wholly practical, and you are encouraged to try out the
examples and exercises for yourself to get hands-on experience. Wherever
possible, small but complete programs and modules are used as examples to
provide realistic use cases. The examples, exercise solutions, and the book’s
errata are available online at
Two sets of examples are provided. The standard examples work with any
Python 3.x version—use these if you care about Python 3.0 compatibility. The
“eg31” examples work with Python 3.1 or later—use these if you don’t need to
support Python 3.0 because your programs’ users have Python 3.1 or later. All
of the examples have been tested on Windows, Linux, and Mac OS X.
While it is best to use the most recent version of Python 3, this is not always
possible if your users cannot or will not upgrade. Every example in this book
works with Python 3.0 except where stated, and those examples and features
that are specific to Python 3.1 are clearly indicated as such.
Although it is possible to use this book to develop software that uses only
Python 3.0, for those wanting to produce software that is expected to be in use
for many years and that is expected to be compatible with later Python 3.x releases, it is best to use Python 3.1 as the oldest Python 3 version that you support. This is partly because Python 3.1 has some very nice new features, but
mostly because the Python developers strongly recommend using Python 3.1
(or later). The developers have decided that Python 3.0.1 will be the last
Python 3.0.y release, and that there will be no more Python 3.0.y releases even
if bugs or security problems are discovered. Instead, they want all Python 3
users to migrate to Python 3.1 (or to a later version), which will have the usual bugfix and security maintenance releases that Python versions normally have.
The Structure of the Book
Chapter 1 presents eight key pieces of Python that are sufficient for writing
complete programs. It also describes some of the Python programming
environments that are available and presents two tiny example programs, both
built using the eight key pieces of Python covered earlier in the chapter.
Chapters 2 through 5 introduce Python’s procedural programming features,
including its basic data types and collection data types, and many useful builtin functions and control structures, as well as very simple text file handling.
Chapter 5 shows how to create custom modules and packages and provides an
overview of Python’s standard library so that you will have a good idea of the
functionality that Python provides out of the box and can avoid reinventing
the wheel.
Chapter 6 provides a thorough introduction to object-oriented programming
with Python. All of the material on procedural programming that you learned
in earlier chapters is still applicable, since object-oriented programming is
built on procedural foundations—for example, making use of the same data
types, collection data types, and control structures.
Chapter 7 covers writing and reading files. For binary files, the coverage includes compression and random access, and for text files, the coverage includes
parsing manually and with regular expressions. This chapter also shows how
to write and read XML files, including using element trees, DOM (Document
Object Model), and SAX (Simple API for XML).
Chapter 8 revisits material covered in some earlier chapters, exploring many of
Python’s more advanced features in the areas of data types and collection data
types, control structures, functions, and object-oriented programming. This
chapter also introduces many new functions, classes, and advanced techniques,
including functional-style programming and the use of coroutines—the material it covers is both challenging and rewarding.
Chapter 9 is different from all the other chapters in that it discusses techniques
and libraries for debugging, testing, and profiling programs, rather than
introducing new Python features.
The remaining chapters cover various advanced topics. Chapter 10 shows techniques for spreading a program’s workload over multiple processes and over
multiple threads. Chapter 11 shows how to write client/server applications
using Python’s standard networking support. Chapter 12 covers database programming (both simple key–value “DBM” files and SQL databases).
Chapter 13 explains and illustrates Python’s regular expression mini-language
and covers the regular expressions module. Chapter 14 follows on from the regular expressions chapter by showing basic parsing techniques using regular expressions, and also using two third-party modules, PyParsing and PLY. Finally,
Chapter 15 introduces GUI (Graphical User Interface) programming using the
tkinter module that is part of Python’s standard library. In addition, the book
has a very brief epilogue, a selected bibliography, and of course, an index.
Most of the book’s chapters are quite long to keep all the related material
together in one place for ease of reference. However, the chapters are broken
down into sections, subsections, and sometimes subsubsections, so it is easy to
read at a pace that suits you; for example, by reading one section or subsection
at a time.
Obtaining and Installing Python 3
If you have a modern and up-to-date Mac or other Unix-like system you may
already have Python 3 installed. You can check by typing python -V (note the
capital V) in a console ( on Mac OS X)—if the version is 3.x you’ve
already got Python 3 and don’t have to install it yourself. If Python wasn’t
found at all it may be that it has a name which includes a version number. Try
typing python3 -V, and if that does not work try python3.0 -V, and failing that try
python3.1 -V. If any of these work you now know that you already have Python
installed, what version it is, and what it is called. (In this book we use the name
python3, but use whatever name worked for you, for example, python3.1.) If you
don’t have any version of Python 3 installed, read on.
For Windows and Mac OS X, easy-to-use graphical installer packages are provided that take you step-by-step through the installation process. These are
available from For Windows, download the “Windows
x86 MSI Installer”, unless you know for sure that your machine has a different
processor for which a separate installer is supplied—for example, if you have
an AMD64, get the “Windows AMD64 MSI Installer”. Once you’ve got the installer, just run it and follow the on-screen instructions.
For Linux, BSD, and other Unixes (apart from Mac OS X for which a .dmg installation file is provided), the easiest way to install Python is to use your operating system’s package management system. In most cases Python is provided
in several separate packages. For example, in Ubuntu (from version 8), there
is python3.0 for Python, idle-python3.0 for IDLE (a simple development environment), and python3.0-doc for the documentation—as well as many other
packages that provide add-ons for even more functionality than that provided
by the standard library. (Naturally, the package names will start with python3.1 for the Python 3.1 versions, and so on.)
If no Python 3 packages are available for your operating system you will
need to download the source from and build Python
from scratch. Get either of the source tarballs and unpack it using tar xvfz
Python-3.1.tgz if you got the gzipped tarball or tar xvfj Python-3.1.tar.bz2 if
you got the bzip2 tarball. (The version numbers may be different, for example,
Python-3.1.1.tgz or Python-3.1.2.tar.bz2, in which case simply replace 3.1 with
your actual version number throughout.) The configuration and building are
standard. First, change into the newly created Python-3.1 directory and run
./configure. (You can use the --prefix option if you want to do a local install.)
Next, run make.
It is possible that you may get some messages at the end saying that not all
modules could be built. This normally means that you don’t have some of the
required libraries or headers on your machine. For example, if the readline
module could not be built, use the package management system to install the
corresponding development library; for example, readline-devel on Fedorabased systems and readline-dev on Debian-based systems such as Ubuntu.
Another module that may not build straight away is the tkinter module—this
depends on both the Tcl and Tk development libraries, tcl-devel and tk-devel
on Fedora-based systems, and tcl8.5-dev and tk8.5-dev on Debian-based systems (and where the minor version may not be 5). Unfortunately, the relevant
package names are not always so obvious, so you might need to ask for help on
Python’s mailing list. Once the missing packages are installed, run ./configure
and make again.
After successfully making, you could run make test to see that everything is
okay, although this is not necessary and can take many minutes to complete.
If you used --prefix to do a local installation, just run make install. For
Python 3.1, if you installed into, say, ~/local/python31, then by adding the ~/local/python31/bin directory to your PATH, you will be able to run Python using
python3 and IDLE using idle3. Alternatively, if you already have a local directory for executables that is already in your PATH (such as ~/bin), you might prefer
to add soft links instead of changing the PATH. For example, if you keep executables in ~/bin and you installed Python in ~/local/python31, you could create
suitable links by executing ln -s ~/local/python31/bin/python3 ~/bin/python3,
and ~/local/python31/bin/idle3 ~/bin/idle3. For this book we did a local install
and added soft links on Linux and Mac OS X exactly as described here—and
on Windows we used the binary installer.
If you did not use --prefix and have root access, log in as root and do make install. On sudo-based systems like Ubuntu, do sudo make install. If Python 2 is
on the system, /usr/bin/python won’t be changed, and Python 3 will be available as python3.0 (or python3.1 depending on the version installed) and from
Python 3.1, in addition, as python3. Python 3.0’s IDLE is installed as idle,
so if access to Python 2’s IDLE is still required the old IDLE will need to be
renamed—for example, to /usr/bin/idle2—before doing the install. Python 3.1
installs IDLE as idle3 and so does not conflict with Python 2’s IDLE.
I would first like to acknowledge with thanks the feedback I have received
from readers of the first edition, who gave corrections, or made suggestions,
or both.
My next acknowledgments are of the book’s technical reviewers, starting
with Jasmin Blanchette, a computer scientist, programmer, and writer with
whom I have cowritten two C++/Qt books. Jasmin’s involvement with chapter
planning and his suggestions and criticisms regarding all the examples, as well
as his careful reading, have immensely improved the quality of this book.
Georg Brandl is a leading Python developer and documentor responsible for
creating Python’s new documentation tool chain. Georg spotted many subtle mistakes and very patiently and persistently explained them until they
were understood and corrected. He also made many improvements to the examples.
Phil Thompson is a Python expert and the creator of PyQt, probably the best
Python GUI library available. Phil’s sharp-eyed and sometimes challenging
feedback led to many clarifications and corrections.
Trenton Schulz is a senior software engineer at Nokia’s Qt Software (formerly
Trolltech) who has been a valuable reviewer of all my previous books, and has
once again come to my aid. Trenton’s careful reading and the numerous suggestions that he made helped clarify many issues and have led to considerable
improvements to the text.
In addition to the aforementioned reviewers, all of whom read the whole
book, David Boddie, a senior technical writer at Nokia’s Qt Software and an
experienced Python practitioner and open source developer, has read and given
valuable feedback on portions of it.
For this second edition, I would also like to thank Paul McGuire (author of the
PyParsing module), who was kind enough to review the PyParsing examples
that appear in the new chapter on parsing, and who gave me a lot of thoughtful
and useful advice. And for the same chapter, David Beazley (author of the
PLY module) reviewed the PLY examples and provided valuable feedback. In
addition, Jasmin, Trenton, Georg, and Phil read most of this second edition’s
new material, and provided very valuable feedback.
Thanks are also due to Guido van Rossum, creator of Python, as well as to the
wider Python community who have contributed so much to make Python, and
especially its libraries, so useful and enjoyable to use.
As always, thanks to Jeff Kingston, creator of the Lout typesetting language
that I have used for more than a decade.
Special thanks to my editor, Debra Williams Cauley, for her support, and for
once again making the entire process as smooth as possible. Thanks also to
Anna Popick, who managed the production process so well, and to the proofreader, Audrey Doyle, who did such fine work once again. And for this second
edition I also want to thank Jennifer Lindner for helping me keep the new material understandable, and the first edition’s Japanese translator Takahiro Na, for spotting some subtle mistakes which I’ve been able to correct
in this edition.
Last but not least, I want to thank my wife, Andrea, both for putting up with
the 4 a.m. wake-ups when book ideas and code corrections often arrived and
insisted upon being noted or tested there and then, and for her love, loyalty,
and support.
● Creating and Running Python
● Python’s “Beautiful Heart”
Rapid Introduction to
Procedural Programming
This chapter provides enough information to get you started writing Python
programs. We strongly recommend that you install Python if you have not
already done so, so that you can get hands-on experience to reinforce what you
learn here. (The Introduction explains how to obtain and install Python on all
major platforms; 4 ➤.)
This chapter’s first section shows you how to create and execute Python programs. You can use your favorite plain text editor to write your Python code,
but the IDLE programming environment discussed in this section provides not
only a code editor, but also additional functionality, including facilities for experimenting with Python code, and for debugging Python programs.
The second section presents eight key pieces of Python that on their own are
sufficient to write useful programs. These pieces are all covered fully in later
chapters, and as the book progresses they are supplemented by all of the rest
of Python so that by the end of the book, you will have covered the whole
language and will be able to use all that it offers in your programs.
The chapter’s final section introduces two short programs which use the subset
of Python features introduced in the second section so that you can get an
immediate taste of Python programming.
Creating and Running Python Programs
Python code can be written using any plain text editor that can load and save
text using either the ASCII or the UTF-8 Unicode character encoding. By default, Python files are assumed to use the UTF-8 character encoding, a superset of ASCII that can represent pretty well every character in every language.
Python files normally have an extension of .py, although on some Unix-like sys9
➤ 91
Chapter 1. Rapid Introduction to Procedural Programming
tems (e.g., Linux and Mac OS X) some Python applications have no extension,
and Python GUI (Graphical User Interface) programs usually have an extension of .pyw, particularly on Windows and Mac OS X. In this book we always use
an extension of .py for Python console programs and Python modules, and .pyw
for GUI programs. All the examples presented in this book run unchanged on
all platforms that have Python 3 available.
Just to make sure that everything is set up correctly, and to show the classical first example, create a file called in a plain text editor (Windows Notepad is fine—we’ll use a better editor shortly), with the following
#!/usr/bin/env python3
print("Hello", "World!")
The first line is a comment. In Python, comments begin with a # and continue to
the end of the line. (We will explain the rather cryptic comment in a moment.)
The second line is blank—outside quoted strings, Python ignores blank lines,
but they are often useful to humans to break up large blocks of code to make
them easier to read. The third line is Python code. Here, the print() function
is called with two arguments, each of type str (string; i.e., a sequence of characters).
Each statement encountered in a .py file is executed in turn, starting with
the first one and progressing line by line. This is different from some other
languages, for example, C++ and Java, which have a particular function or
method with a special name where they start from. The flow of control can of
course be diverted as we will see when we discuss Python’s control structures
in the next section.
We will assume that Windows users keep their Python code in the C:\py3eg
directory and that Unix (i.e., Unix, Linux, and Mac OS X) users keep their code
in the $HOME/py3eg directory. Save into the py3eg directory and close
the text editor.
Now that we have a program, we can run it. Python programs are executed
by the Python interpreter, and normally this is done inside a console window.
On Windows the console is called “Console”, or “DOS Prompt”, or “MS-DOS
Prompt”, or something similar, and is usually available from Start→All Programs→Accessories. On Mac OS X the console is provided by the program (located in Applications/Utilities by default), available using Finder, and
on other Unixes, we can use an xterm or the console provided by the windowing
environment, for example, konsole or gnome-terminal.
Start up a console, and on Windows enter the following commands (which
assume that Python is installed in the default location)—the console’s output
is shown in lightface; what you type is shown in bold:
Creating and Running Python Programs
C:\>cd c:\py3eg
Since the cd (change directory) command has an absolute path, it doesn’t
matter which directory you start out from.
Unix users enter this instead (assuming that Python 3 is in the PATH):★
$ cd $HOME/py3eg
$ python3
In both cases the output should be the same:
Hello World!
Note that unless stated otherwise, Python’s behavior on Mac OS X is the
same as that on any other Unix system. In fact, whenever we refer to “Unix”
it can be taken to mean Linux, BSD, Mac OS X, and most other Unixes and
Unix-like systems.
Although the program has just one executable statement, by running it we can
infer some information about the print() function. For one thing, print() is a
built-in part of the Python language—we didn’t need to “import” or “include”
it from a library to make use of it. Also, it separates each item it prints with
a single space, and prints a newline after the last item is printed. These are
default behaviors that can be changed, as we will see later. Another thing
worth noting about print() is that it can take as many or as few arguments as
we care to give it.
Typing such command lines to invoke our Python programs would quickly
become tedious. Fortunately, on both Windows and Unix we can use more
convenient approaches. Assuming we are in the py3eg directory, on Windows
we can simply type:
Windows uses its registry of file associations to automatically call the Python
interpreter when a filename with extension .py is entered in a console.
Unfortunately, this convenience does not always work, since some versions
of Windows have a bug that sometimes affects the execution of interpreted
programs that are invoked as the result of a file association. This isn’t specific
to Python; other interpreters and even some .bat files are affected by the bug
too. If this problem arises, simply invoke Python directly rather than relying
on the file association.
If the output on Windows is:
The Unix prompt may well be different from the $ shown here; it does not matter what it is.
➤ 181
Chapter 1. Rapid Introduction to Procedural Programming
('Hello', 'World!')
then it means that Python 2 is on the system and is being invoked instead
of Python 3. One solution to this is to change the .py file association from
Python 2 to Python 3. The other (less convenient, but safer) solution is to put
the Python 3 interpreter in the path (assuming it is installed in the default location), and execute it explicitly each time. (This also gets around the Windows
file association bug mentioned earlier.) For example:
It might be more convenient to create a py3.bat file with the single line
path=c:\python31;%path% and to save this file in the C:\Windows directory. Then,
whenever you start a console for running Python 3 programs, begin by executing py3.bat. Or alternatively you can have py3.bat executed automatically.
To do this, change the console’s properties (find the console in the Start menu,
then right-click it to pop up its Properties dialog), and in the Shortcut tab’s Target
string, append the text “ /u /k c:\windows\py3.bat” (note the space before,
between, and after the “/u” and “/k” options, and be sure to add this at the end
after “cmd.exe”).
On Unix, we must first make the file executable, and then we can run it:
$ chmod +x
$ ./
We need to run the chmod command only once of course; after that we can
simply enter ./ and the program will run.
On Unix, when a program is invoked in the console, the file’s first two bytes are
read.★ If these bytes are the ASCII characters #!, the shell assumes that the file
is to be executed by an interpreter and that the file’s first line specifies which
interpreter to use. This line is called the shebang (shell execute) line, and if
present must be the first line in the file.
The shebang line is commonly written in one of two forms, either:
#!/usr/bin/env python3
If written using the first form, the specified interpreter is used. This form
may be necessary for Python programs that are to be run by a web server,
The interaction between the user and the console is handled by a “shell” program. The distinction
between the console and the shell does not concern us here, so we use the terms interchangeably.
Creating and Running Python Programs
although the specific path may be different from the one shown. If written
using the second form, the first python3 interpreter found in the shell’s current
environment is used. The second form is more versatile because it allows for
the possibility that the Python 3 interpreter is not located in /usr/bin (e.g., it
could be in /usr/local/bin or installed under $HOME). The shebang line is not
needed (but is harmless) under Windows; all the examples in this book have a
shebang line of the second form, although we won’t show it.
Note that for Unix systems we assume that the name of Python 3’s executable
(or a soft link to it) in the PATH is python3. If this is not the case, you will need
to change the shebang line in the examples to use the correct name (or correct
name and path if you use the first form), or create a soft link from the Python 3
executable to the name python3 somewhere in the PATH.
Obtaining and
Many powerful plain text editors, such as Vim and Emacs, come with built-in
support for editing Python programs. This support typically involves providing
color syntax highlighting and correctly indenting or unindenting lines. An alternative is to use the IDLE Python programming environment. On Windows
and Mac OS X, IDLE is installed by default. On Unixes IDLE is built along
with the Python interpreter if you build from the tarball, but if you use a package manager, IDLE is usually provided as a separate package as described in
the Introduction.
As the screenshot in Figure 1.1 shows, IDLE has a rather retro look that harks
back to the days of Motif on Unix and Windows 95. This is because it uses the
Tk-based Tkinter GUI library (covered in Chapter 15) rather than one of the
more powerful modern GUI libraries such as PyGtk, PyQt, or wxPython. The
reasons for the use of Tkinter are a mixture of history, liberal license conditions, and the fact that Tkinter is much smaller than the other GUI libraries.
On the plus side, IDLE comes as standard with Python and is very simple to
learn and use.
IDLE provides three key facilities: the ability to enter Python expressions
and code and to see the results directly in the Python Shell; a code editor that
provides Python-specific color syntax highlighting and indentation support;
and a debugger that can be used to step through code to help identify and kill
bugs. The Python Shell is especially useful for trying out simple algorithms,
snippets of code, and regular expressions, and can also be used as a very
powerful and flexible calculator.
Several other Python development environments are available, but we recommend that you use IDLE, at least at first. An alternative is to create your programs in the plain text editor of your choice and debug using calls to print().
It is possible to invoke the Python interpreter without specifying a Python
program. If this is done the interpreter starts up in interactive mode. In
this mode it is possible to enter Python statements and see the results exactly
the same as when using IDLE’s Python Shell window, and with the same >>>
Chapter 1. Rapid Introduction to Procedural Programming
Figure 1.1 IDLE’s Python Shell
prompts. But IDLE is much easier to use, so we recommend using IDLE for
experimenting with code snippets. The short interactive examples we show
are all assumed to be entered in an interactive Python interpreter or in IDLE’s
Python Shell.
We now know how to create and run Python programs, but clearly we won’t get
very far knowing only a single function. In the next section we will considerably increase our Python knowledge. This will make us able to create short but
useful Python programs, something we will do in this chapter’s last section.
Python’s “Beautiful Heart”
In this section we will learn about eight key pieces of Python, and in the next
section we will show how these pieces can be used to write a couple of small but
realistic programs. There is much more to say about all of the things covered
in this section, so if as you read it you feel that Python is missing something
or that things are sometimes done in a long-winded way, peek ahead using the
forward references or using the table of contents or index, and you will almost
certainly find that Python has the feature you want and often has more concise
forms of expression than we show here—and a lot more besides.
Piece #1: Data Types
One fundamental thing that any programming language must be able to do
is represent items of data. Python provides several built-in data types, but
we will concern ourselves with only two of them for now. Python represents
Python’s “Beautiful Heart”
integers (positive and negative whole numbers) using the int type, and it
represents strings (sequences of Unicode characters) using the str type. Here
are some examples of integer and string literals:
"Infinitely Demanding"
'Simon Critchley'
'positively αβγ ÷©'
Incidentally, the second number shown is 2217—the size of Python’s integers
is limited only by machine memory, not by a fixed number of bytes. Strings
can be delimited by double or single quotes, as long as the same kind are used
at both ends, and since Python uses Unicode, strings are not limited to ASCII
characters, as the penultimate string shows. An empty string is simply one
with nothing between the delimiters.
Python uses square brackets ([]) to access an item from a sequence such as
a string. For example, if we are in a Python Shell (either in the interactive
interpreter, or in IDLE) we can enter the following—the Python Shell’s output
is shown in lightface; what you type is shown in bold:
>>> "Hard Times"[5]
>>> "giraffe"[0]
Traditionally, Python Shells use >>> as their prompt, although this can be
changed. The square brackets syntax can be used with data items of any data
type that is a sequence, such as strings and lists. This consistency of syntax
is one of the reasons that Python is so beautiful. Note that all Python index
positions start at 0.
In Python, both str and the basic numeric types such as int are immutable—that is, once set, their value cannot be changed. At first this appears
to be a rather strange limitation, but Python’s syntax means that this is a nonissue in practice. The only reason for mentioning it is that although we can use
square brackets to retrieve the character at a given index position in a string,
we cannot use them to set a new character. (Note that in Python a character is
simply a string of length 1.)
To convert a data item from one type to another we can use the syntax
datatype(item). For example:
>>> int("45")
Chapter 1. Rapid Introduction to Procedural Programming
>>> str(912)
The int() conversion is tolerant of leading and trailing whitespace, so
int(" 45 ") would have worked just as well. The str() conversion can be
applied to almost any data item. We can easily make our own custom data
types support str() conversion, and also int() or other conversions if they
make sense, as we will see in Chapter 6. If a conversion fails, an exception is
raised—we briefly introduce exception-handling in Piece #5, and fully cover
exceptions in Chapter 4.
Strings and integers are fully covered in Chapter 2, along with other built-in
data types and some data types from Python’s standard library. That chapter
also covers operations that can be applied to immutable sequences, such
as strings.
Piece #2: Object References
Once we have some data types, the next thing we need are variables in which
to store them. Python doesn’t have variables as such, but instead has object
references. When it comes to immutable objects like ints and strs, there is
no discernable difference between a variable and an object reference. As for
mutable objects, there is a difference, but it rarely matters in practice. We will
use the terms variable and object reference interchangeably.
Let’s look at a few tiny examples, and then discuss some of the details.
x = "blue"
y = "green"
z = x
The syntax is simply objectReference = value. There is no need for predeclaration and no need to specify the value’s type. When Python executes the first
statement it creates a str object with the text “blue”, and creates an object reference called x that refers to the str object. For all practical purposes we can
say that “variable x has been assigned the ‘blue’ string”. The second statement
is similar. The third statement creates a new object reference called z and sets
it to refer to the same object that the x object reference refers to (in this case
the str containing the text “blue”).
The = operator is not the same as the variable assignment operator in some
other languages. The = operator binds an object reference to an object in
memory. If the object reference already exists, it is simply re-bound to refer to
the object on the right of the = operator; if the object reference does not exist it
is created by the = operator.
➤ 146
Chapter 1. Rapid Introduction to Procedural Programming
route = "North"
print(route, type(route)) # prints: North <class 'str'>
Here we create a new object reference called route and set it to refer to a new
int of value 866. At this point we could use / with route since division is a valid
operation for integers. Then we reuse the route object reference to refer to a
new str of value “North”, and the int object is scheduled for garbage collection
since now no object reference refers to it. At this point using / with route would
cause a TypeError to be raised since / is not a valid operation for a string.
The type() function returns the data type (also known as the “class”) of the
data item it is given—this function can be very useful for testing and debugging, but would not normally appear in production code, since there is a better
alternative as we will see in Chapter 6.
If we are experimenting with Python code inside the interactive interpreter or
in a Python Shell such as the one provided by IDLE, simply typing the name
of an object reference is enough to have Python print its value. For example:
>>> x = "blue"
>>> y = "green"
>>> z = x
>>> x
>>> x, y, z
('blue', 'green', 'blue')
This is much more convenient than having to call the print() function all
the time, but works only when using Python interactively—any programs
and modules that we write must use print() or similar functions to produce
output. Notice that Python displayed the last output in parentheses separated
by commas—this signifies a tuple, that is, an ordered immutable sequence of
objects. We will cover tuples in the next piece.
Piece #3: Collection Data Types
It is often convenient to hold entire collections of data items. Python provides
several collection data types that can hold items, including associative arrays
and sets. But here we will introduce just two: tuple and list. Python tuples and
lists can be used to hold any number of data items of any data types. Tuples
are immutable, so once they are created we cannot change them. Lists are
mutable, so we can easily insert items and remove items whenever we want.
Tuples are created using commas (,), as these examples show—and note that
here, and from now on, we don’t use bold to distinguish what you type:
>>> "Denmark", "Finland", "Norway", "Sweden"
('Denmark', 'Finland', 'Norway', 'Sweden')
➤ 242
Python’s “Beautiful Heart”
>>> "one",
When Python outputs a tuple it encloses it in parentheses. Many programmers
emulate this and always enclose the tuple literals they write in parentheses.
If we have a one-item tuple and want to use parentheses, we must still use
the comma—for example, (1,). An empty tuple is created by using empty
parentheses, (). The comma is also used to separate arguments in function
calls, so if we want to pass a tuple literal as an argument we must enclose it in
parentheses to avoid confusion.
Here are some example lists:
➤ 36
➤ 108
Creating and
[1, 4, 9, 16, 25, 36, 49]
['alpha', 'bravo', 'charlie', 'delta', 'echo']
['zebra', 49, -879, 'aardvark', 200]
One way to create a list is to use square brackets ([]) as we have done here;
later on we will see other ways. The fourth list shown is an empty list.
Under the hood, lists and tuples don’t store data items at all, but rather object
references. When lists and tuples are created (and when items are inserted in
the case of lists), they take copies of the object references they are given. In
the case of literal items such as integers or strings, an object of the appropriate
data type is created in memory and suitably initialized, and then an object
reference referring to the object is created, and it is this object reference that
is put in the list or tuple.
Like everything else in Python, collection data types are objects, so we can nest
collection data types inside other collection data types, for example, to create
lists of lists, without formality. In some situations the fact that lists, tuples,
and most of Python’s other collection data types hold object references rather
than objects makes a difference—this is covered in Chapter 3.
In procedural programming we call functions and often pass in data items as
arguments. For example, we have already seen the print() function. Another
frequently used Python function is len(), which takes a single data item as its
argument and returns the “length” of the item as an int. Here are a few calls
to len():
>>> len(("one",))
>>> len([3, 5, 1, 2, "pause", 5])
>>> len("automatically")
➤ 113
➤ 146
Chapter 1. Rapid Introduction to Procedural Programming
Tuples, lists, and strings are “sized”, that is, they are data types that have
a notion of size, and data items of any such data type can be meaningfully
passed to the len() function. (An exception is raised if a nonsized data item is
passed to len().)
All Python data items are objects (also called instances) of a particular data
type (also called a class). We will use the terms data type and class interchangeably. One key difference between an object, and the plain items of data that
some other languages provide (e.g., C++ or Java’s built-in numeric types), is
that an object can have methods. Essentially, a method is simply a function
that is called for a particular object. For example, the list type has an append()
method, so we can append an object to a list like this:
>>> x = ["zebra", 49, -879, "aardvark", 200]
>>> x.append("more")
>>> x
['zebra', 49, -879, 'aardvark', 200, 'more']
The x object knows that it is a list (all Python objects know what their own
data type is), so we don’t need to specify the data type explicitly. In the implementation of the append() method the first argument will be the x object
itself—this is done automatically by Python as part of its syntactic support for
The append() method mutates, that is, changes, the original list. This is possible because lists are mutable. It is also potentially more efficient than creating
a new list with the original items and the extra item and then rebinding the
object reference to the new list, particularly for very long lists.
In a procedural language the same thing could be achieved by using the list’s
append() like this (which is perfectly valid Python syntax):
>>> list.append(x, "extra")
>>> x
['zebra', 49, -879, 'aardvark', 200, 'more', 'extra']
Here we specify the data type and the data type’s method, and give as the
first argument the data item of the data type we want to call the method on,
followed by any additional arguments. (In the face of inheritance there is a
subtle semantic difference between the two syntaxes; the first form is the one
that is most commonly used in practice. Inheritance is covered in Chapter 6.)
If you are unfamiliar with object-oriented programming this may seem a bit
strange at first. For now, just accept that Python has conventional functions
called like this: functionName(arguments); and methods which are called like
this: objectName.methodName(arguments). (Object-oriented programming is covered in Chapter 6.)
➤ 383
Python’s “Beautiful Heart”
The dot (“access attribute”) operator is used to access an object’s attributes.
An attribute can be any kind of object, although so far we have shown only
method attributes. Since an attribute can be an object that has attributes,
which in turn can have attributes, and so on, we can use as many dot operators
as necessary to access the particular attribute we want.
The list type has many other methods, including insert() which is used to
insert an item at a given index position, and remove() which removes an item at
a given index position. As noted earlier, Python indexes are always 0-based.
We saw before that we can get characters from strings using the square
brackets operator, and noted at the time that this operator could be used with
any sequence. Lists are sequences, so we can do things like this:
>>> x
['zebra', 49, -879, 'aardvark', 200, 'more', 'extra']
>>> x[0]
>>> x[4]
Tuples are also sequences, so if x had been a tuple we could retrieve items using square brackets in exactly the same way as we have done for the x list. But
since lists are mutable (unlike strings and tuples which are immutable), we can
also use the square brackets operator to set list elements. For example:
>>> x[1] = "forty nine"
>>> x
['zebra', 'forty nine', -879, 'aardvark', 200, 'more', 'extra']
If we give an index position that is out of range, an exception will be raised—we
briefly introduce exception-handling in Piece #5, and fully cover exceptions in
Chapter 4.
We have used the term sequence a few times now, relying on an informal understanding of its meaning, and will continue to do so for the time being. However,
Python defines precisely what features a sequence must support, and similarly
defines what features a sized object must support, and so on for various other
categories that a data type might belong to, as we will see in Chapter 8.
Lists, tuples, and Python’s other built-in collection data types are covered in
Chapter 3.
Piece #4: Logical Operations
One of the fundamental features of any programming language is its logical
operations. Python provides four sets of logical operations, and we will review
the fundamentals of all of them here.
Chapter 1. Rapid Introduction to Procedural Programming
The Identity Operator
Since all Python variables are really object references, it sometimes makes
sense to ask whether two or more object references are referring to the same
object. The is operator is a binary operator that returns True if its left-hand object reference is referring to the same object as its right-hand object reference.
Here are some examples:
>>> a
>>> b
>>> a
>>> b
>>> a
= ["Retention", 3, None]
= ["Retention", 3, None]
is b
= a
is b
Note that it usually does not make sense to use is for comparing ints, strs, and
most other data types since we almost invariably want to compare their values.
In fact, using is to compare data items can lead to unintuitive results, as we
can see in the preceding example, where although a and b are initially set to
the same list values, the lists themselves are held as separate list objects and
so is returns False the first time we use it.
One benefit of identity comparisons is that they are very fast. This is because
the objects referred to do not have to be examined themselves. The is operator
needs to compare only the memory addresses of the objects—the same address
means the same object.
The most common use case for is is to compare a data item with the built-in
null object, None, which is often used as a place-marking value to signify
“unknown” or “nonexistent”:
>>> a = "Something"
>>> b = None
>>> a is not None, b is None
(True, True)
To invert the identity test we use is not.
The purpose of the identity operator is to see whether two object references
refer to the same object, or to see whether an object is None. If we want to
compare object values we should use a comparison operator instead.
Comparison Operators
Python provides the standard set of binary comparison operators, with the
expected semantics: < less than, <= less than or equal to, == equal to, != not
Python’s “Beautiful Heart”
equal to, >= greater than or equal to, and > greater than. These operators
compare object values, that is, the objects that the object references used in the
comparison refer to. Here are a few examples typed into a Python Shell:
>>> a = 2
>>> b = 6
>>> a == b
>>> a < b
>>> a <= b, a != b, a >= b, a > b
(True, True, False, False)
Everything is as we would expect with integers. Similarly, strings appear to
compare properly too:
>>> a
>>> b
>>> a
>>> a
= "many paths"
= "many paths"
is b
== b
Although a and b are different objects (have different identities), they have
the same values, so they compare equal. Be aware, though, that because
Python uses Unicode for representing strings, comparing strings that contain
non-ASCII characters can be a lot subtler and more complicated than it might
at first appear—we will fully discuss this issue in Chapter 2.
In some cases, comparing the identity of two strings or numbers—for example,
using a is b—will return True, even if each has been assigned separately as we
did here. This is because some implementations of Python will reuse the same
object (since the value is the same and is immutable) for the sake of efficiency.
The moral of this is to use == and != when comparing values, and to use is and
is not only when comparing with None or when we really do want to see if two
object references, rather than their values, are the same.
One particularly nice feature of Python’s comparison operators is that they can
be chained. For example:
>>> a = 9
>>> 0 <= a <= 10
This is a nicer way of testing that a given data item is in range than having
to do two separate comparisons joined by logical and, as most other languages
require. It also has the additional virtue of evaluating the data item only once
(since it appears once only in the expression), something that could make a
➤ 68
Chapter 1. Rapid Introduction to Procedural Programming
difference if computing the data item’s value is expensive, or if accessing the
data item causes side effects.
Thanks to the “strong” aspect of Python’s dynamic typing, comparisons that
don’t make sense will cause an exception to be raised. For example:
>>> "three" < 4
Traceback (most recent call last):
TypeError: unorderable types: str() < int()
When an exception is raised and not handled, Python outputs a traceback
along with the exception’s error message. For clarity, we have omitted the
traceback part of the output, replacing it with an ellipsis.★ The same TypeError
exception would occur if we wrote "3" < 4 because Python does not try to guess
our intentions—the right approach is either to explicitly convert, for example,
int("3") < 4, or to use comparable types, that is, both integers or both strings.
Python makes it easy for us to create custom data types that will integrate
nicely so that, for example, we could create our own custom numeric type
which would be able to participate in comparisons with the built-in int type,
and with other built-in or custom numeric types, but not with strings or other
non-numeric types.
The Membership Operator
For data types that are sequences or collections such as strings, lists, and tuples, we can test for membership using the in operator, and for nonmembership
using the not in operator. For example:
>>> p = (4, "frog", 9, -33, 9, 2)
>>> 2 in p
>>> "dog" not in p
For lists and tuples, the in operator uses a linear search which can be slow for
very large collections (tens of thousands of items or more). On the other hand,
in is very fast when used on a dictionary or a set; both of these collection data
types are covered in Chapter 3. Here is how in can be used with a string:
>>> phrase = "Wild Swans by Jung Chang"
>>> "J" in phrase
A traceback (sometimes called a backtrace) is a list of all the calls made from the point where the
unhandled exception occurred back to the top of the call stack.
➤ 415
➤ 256
Python’s “Beautiful Heart”
>>> "han" in phrase
Conveniently, in the case of strings, the membership operator can be used to
test for substrings of any length. (As noted earlier, a character is just a string
of length 1.)
Logical Operators
Python provides three logical operators: and, or, and not. Both and and or use
short-circuit logic and return the operand that determined the result—they do
not return a Boolean (unless they actually have Boolean operands). Let’s see
what this means in practice:
five = 5
two = 2
zero = 0
five and two
two and five
five and zero
If the expression occurs in a Boolean context, the result is evaluated as a
Boolean, so the preceding expressions would come out as True, True, and False
in, say, an if statement.
nought = 0
five or two
two or five
zero or five
zero or nought
The or operator is similar; here the results in a Boolean context would be True,
True, True, and False.
The not unary operator evaluates its argument in a Boolean context and
always returns a Boolean result, so to continue the earlier example, not
(zero or nought) would produce True, and not two would produce False.
Chapter 1. Rapid Introduction to Procedural Programming
Piece #5: Control Flow Statements
We mentioned earlier that each statement encountered in a .py file is executed
in turn, starting with the first one and progressing line by line. The flow of
control can be diverted by a function or method call or by a control structure,
such as a conditional branch or a loop statement. Control is also diverted when
an exception is raised.
In this subsection we will look at Python’s if statement and its while and for
loops, deferring consideration of functions to Piece #8, and methods to Chapter 6. We will also look at the very basics of exception-handling; we cover the
subject fully in Chapter 4. But first we will clarify a couple of items of terminology.
A Boolean expression is anything that can be evaluated to produce a Boolean
value (True or False). In Python, such an expression evaluates to False if it is
the predefined constant False, the special object None, an empty sequence or
collection (e.g., an empty string, list, or tuple), or a numeric data item of value
0; anything else is considered to be True. When we create our own custom data
types (e.g., in Chapter 6), we can decide for ourselves what they should return
in a Boolean context.
In Python-speak a block of code, that is, a sequence of one or more statements,
is called a suite. Because some of Python’s syntax requires that a suite be
present, Python provides the keyword pass which is a statement that does
nothing and that can be used where a suite is required (or where we want to
indicate that we have considered a particular case) but where no processing
is necessary.
The if Statement
The general syntax for Python’s if statement is this:★
if boolean_expression1:
elif boolean_expression2:
elif boolean_expressionN:
In this book, ellipses (…) are used to indicate lines that are not shown.
Python’s “Beautiful Heart”
There can be zero or more elif clauses, and the final else clause is optional. If
we want to account for a particular case, but want to do nothing if it occurs, we
can use pass as that branch’s suite.
The first thing that stands out to programmers used to C++ or Java is that
there are no parentheses and no braces. The other thing to notice is the
colon: This is part of the syntax and is easy to forget at first. Colons are used
with else, elif, and essentially in any other place where a suite is to follow.
Unlike most other programming languages, Python uses indentation to signify
its block structure. Some programmers don’t like this, especially before they
have tried it, and some get quite emotional about the issue. But it takes just a
few days to get used to, and after a few weeks or months, brace-free code seems
much nicer and less cluttered to read than code that uses braces.
Since suites are indicated using indentation, the question that naturally arises is, “What kind of indentation?” The Python style guidelines recommend
four spaces per level of indentation, and only spaces (no tabs). Most modern
text editors can be set up to handle this automatically (IDLE’s editor does of
course, and so do most other Python-aware editors). Python will work fine with
any number of spaces or with tabs or with a mixture of both, providing that
the indentation used is consistent. In this book, we follow the official Python
Here is a very simple if statement example:
if x:
print("x is nonzero")
In this case, if the condition (x) evaluates to True, the suite (the print() function
call) will be executed.
if lines < 1000:
elif lines < 10000:
This is a slightly more elaborate if statement that prints a word that describes
the value of the lines variable.
The while Statement
The while statement is used to execute a suite zero or more times, the number
of times depending on the state of the while loop’s Boolean expression. Here’s
the syntax:
Chapter 1. Rapid Introduction to Procedural Programming
while boolean_expression:
Actually, the while loop’s full syntax is more sophisticated than this, since both
break and continue are supported, and also an optional else clause that we will
discuss in Chapter 4. The break statement switches control to the statement
following the innermost loop in which the break statement appears—that is,
it breaks out of the loop. The continue statement switches control to the start
of the loop. Both break and continue are normally used inside if statements to
conditionally change a loop’s behavior.
while True:
item = get_next_item()
if not item:
This while loop has a very typical structure and runs as long as there are items
to process. (Both get_next_item() and process_item() are assumed to be custom
functions defined elsewhere.) In this example, the while statement’s suite
contains an if statement, which itself has a suite—as it must—in this case
consisting of a single break statement.
The for … in Statement
Python’s for loop reuses the in keyword (which in other contexts is the membership operator), and has the following syntax:
for variable in iterable:
Just like the while loop, the for loop supports both break and continue, and also
has an optional else clause. The variable is set to refer to each object in the
iterable in turn. An iterable is any data type that can be iterated over, and
includes strings (where the iteration is character by character), lists, tuples,
and Python’s other collection data types.
for country in ["Denmark", "Finland", "Norway", "Sweden"]:
Here we take a very simplistic approach to printing a list of countries. In
practice it is much more common to use a variable:
countries = ["Denmark", "Finland", "Norway", "Sweden"]
for country in countries:
Python’s “Beautiful Heart”
In fact, an entire list (or tuple) can be printed using the print() function
directly, for example, print(countries), but we often prefer to print collections
using a for loop (or a list comprehension, covered later), to achieve full control
over the formatting.
if letter in "AEIOU":
print(letter, "is a vowel")
print(letter, "is a consonant")
In this snippet the first use of the in keyword is part of a for statement, with
the variable letter taking on the values "A", "B", and so on up to "Z", changing
at each iteration of the loop. On the snippet’s second line we use in again, but
this time as the membership testing operator. Notice also that this example
shows nested suites. The for loop’s suite is the if … else statement, and both
the if and the else branches have their own suites.
Basic Exception Handling
Many of Python’s functions and methods indicate errors or other important
events by raising an exception. An exception is an object like any other Python
object, and when converted to a string (e.g., when printed), the exception
produces a message text. A simple form of the syntax for exception handlers
is this:
except exception1 as variable1:
except exceptionN as variableN:
Note that the as variable part is optional; we may care only that a particular
exception was raised and not be interested in its message text.
The full syntax is more sophisticated; for example, each except clause can
handle multiple exceptions, and there is an optional else clause. All of this is
covered in Chapter 4.
The logic works like this. If the statements in the try block’s suite all execute
without raising an exception, the except blocks are skipped. If an exception
is raised inside the try block, control is immediately passed to the suite corresponding to the first matching exception—this means that any statements in
the suite that follow the one that caused the exception will not be executed. If
➤ 118
Chapter 1. Rapid Introduction to Procedural Programming
this occurs and if the as variable part is given, then inside the exception-handling suite, variable refers to the exception object.
If an exception occurs in the handling except block, or if an exception is raised
that does not match any of the except blocks in the first place, Python looks for
a matching except block in the next enclosing scope. The search for a suitable
exception handler works outward in scope and up the call stack until either
a match is found and the exception is handled, or no match is found, in which
case the program terminates with an unhandled exception. In the case of
an unhandled exception, Python prints a traceback as well as the exception’s
message text.
Here is an example:
s = input("enter an integer: ")
i = int(s)
print("valid integer entered:", i)
except ValueError as err:
If the user enters “3.5”, the output will be:
invalid literal for int() with base 10: '3.5'
But if they were to enter “13”, the output will be:
valid integer entered: 13
Many books consider exception-handling to be an advanced topic and defer it
as late as possible. But raising and especially handling exceptions is fundamental to the way Python works, so we make use of it from the beginning. And as
we shall see, using exception handlers can make code much more readable, by
separating the “exceptional” cases from the processing we are really interested in.
Piece #6: Arithmetic Operators
Python provides a full set of arithmetic operators, including binary operators
for the four basic mathematical operations: + addition, - subtraction, * multiplication, and / division. In addition, many Python data types can be used with
augmented assignment operators such as += and *=. The +, -, and * operators
all behave as expected when both of their operands are integers:
>>> 5 + 6
Dealing with
➤ 415
Python’s “Beautiful Heart”
>>> 3 - 7
>>> 4 * 8
Notice that - can be used both as a unary operator (negation) and as a binary
operator (subtraction), as is common in most programming languages. Where
Python differs from the crowd is when it comes to division:
>>> 12 / 3
>>> 3 / 2
The division operator produces a floating-point value, not an integer; many
other languages will produce an integer, truncating any fractional part. If
we need an integer result, we can always convert using int() (or use the
truncating division operator //, discussed later).
Numeric operators and
➤ 55
a = 5
a += 8
At first sight the preceding statements are unsurprising, particularly to those
familiar with C-like languages. In such languages, augmented assignment is
shorthand for assigning the results of an operation—for example, a += 8 is essentially the same as a = a + 8. However, there are two important subtleties here,
one Python-specific and one to do with augmented operators in any language.
The first point to remember is that the int data type is immutable—that is,
once assigned, an int’s value cannot be changed. So, what actually happens
behind the scenes when an augmented assignment operator is used on an
immutable object is that the operation is performed, and an object holding the
result is created; and then the target object reference is re-bound to refer to the
result object rather than the object it referred to before. So, in the preceding
case when the statement a += 8 is encountered, Python computes a + 8, stores
the result in a new int object, and then rebinds a to refer to this new int. (And
if the original object a was referring to has no more object references referring
to it, it will be scheduled for garbage collection.) Figure 1.3 illustrates this
The second subtlety is that a operator= b is not quite the same as a = a operator
b. The augmented version looks up a’s value only once, so it is potentially faster.
Also, if a is a complex expression (e.g., a list element with an index position
calculation such as items[offset + index]), using the augmented version may
Python’s “Beautiful Heart”
er hand, mutable types can be more convenient to use. Where the distinction
matters, we will discuss it—for example, in Chapter 4 when we discuss setting
default arguments for custom functions, in Chapter 3 when we discuss lists,
sets, and some other data types, and again in Chapter 6 when we show how to
create custom data types.
The right-hand operand for the list += operator must be an iterable; if it is not
an exception is raised:
>>> seeds += 5
Traceback (most recent call last):
TypeError: 'int' object is not iterable
➤ 415
The correct way to extend a list is to use an iterable object, such as a list:
>>> seeds += [5]
>>> seeds
['sesame', 'sunflower', 'pumpkin', 5]
And of course, the iterable object used to extend the list can itself have more
than one item:
>>> seeds += [9, 1, 5, "poppy"]
>>> seeds
['sesame', 'sunflower', 'pumpkin', 5, 9, 1, 5, 'poppy']
Appending a plain string—for example, "durian"—rather than a list containing
a string, ["durian"], leads to a logical but perhaps surprising result:
>>> seeds = ["sesame", "sunflower", "pumpkin"]
>>> seeds += "durian"
>>> seeds
['sesame', 'sunflower', 'pumpkin', 'd', 'u', 'r', 'i', 'a', 'n']
The list += operator extends the list by appending each item of the iterable it
is provided with; and since a string is an iterable, this leads to each character
in the string being appended individually. If we use the list append() method,
the argument is always added as a single item.
Piece #7: Input/Output
To be able to write genuinely useful programs we must be able to read
input—for example, from the user at the console, and from files—and produce
output, either to the console or to files. We have already made use of Python’s
built-in print() function, although we will defer covering it further until Chap-
Dealing with
Chapter 1. Rapid Introduction to Procedural Programming
ter 4. In this subsection we will concentrate on console I/O, and use shell redirection for reading and writing files.
Python provides the built-in input() function to accept input from the user.
This function takes an optional string argument (which it prints on the console); it then waits for the user to type in a response and to finish by pressing
Enter (or Return). If the user does not type any text but just presses Enter, the input() function returns an empty string; otherwise, it returns a string containing what the user typed, without any line terminator.
Here is our first complete “useful” program; it draws on many of the previous
pieces—the only new thing it shows is the input() function:
print("Type integers, each followed by Enter; or just Enter to finish")
total = 0
count = 0
while True:
line = input("integer: ")
if line:
number = int(line)
except ValueError as err:
total += number
count += 1
if count:
print("count =", count, "total =", total, "mean =", total / count)
The program (in file in the book’s examples) has just 17 executable
lines. Here is what a typical run looks like:
Type integers, each followed by Enter; or just Enter to finish
number: 12
number: 7
number: 1x
invalid literal for int() with base 10: '1x'
number: 15
number: 5
count = 4 total = 39 mean = 9.75
Although the program is very short, it is fairly robust. If the user enters a
string that cannot be converted to an integer, the problem is caught by an
Python’s “Beautiful Heart”
exception handler that prints a suitable message and switches control to the
start of the loop (“continues the loop”). And the last if statement ensures that
if the user doesn’t enter any numbers at all, the summary isn’t output, and
division by zero is avoided.
File handling is fully covered in Chapter 7; but right now we can create files by
redirecting the print() functions’ output from the console. For example:
C:\> > results.txt
Windows file
association bug
11 ➤
will cause the output of plain print() function calls made in the fictitious program to be written to the file results.txt. This syntax works in the
Windows console (usually) and in Unix consoles. For Windows, we must write
C:\Python31\python.exe > results.txt if Python 2 is the machine’s default Python version or if the console exhibits the file association bug; otherwise, assuming Python 3 is in the PATH, python > results.txt should be
sufficient, if plain > results.txt doesn’t work. For Unixes we must
make the program executable (chmod +x and then invoke it by typing
./ unless the directory it is in happens to be in the PATH, in which case
invoking it by typing is sufficient.
Reading data can be achieved by redirecting a file of data as input in an
analogous way to redirecting output. However, if we used redirection with, the program would fail. This is because the input() function raises an
exception if it receives an EOF (end of file) character. Here is a more robust
version ( that can accept input from the user typing at the keyboard, or
via file redirection:
print("Type integers, each followed by Enter; or ^D or ^Z to finish")
total = 0
count = 0
while True:
line = input()
if line:
number = int(line)
total += number
count += 1
except ValueError as err:
except EOFError:
if count:
print("count =", count, "total =", total, "mean =", total / count)
Chapter 1. Rapid Introduction to Procedural Programming
Given the command line < data\sum2.dat (where the sum2.dat file contains a list of numbers one per line and is in the examples’ data subdirectory),
the output to the console is:
Type integers, each followed by Enter; or ^D or ^Z to finish
count = 37 total = 1839 mean = 49.7027027027
We have made several small changes to make the program more suitable for
use both interactively and using redirection. First, we have changed the
termination from being a blank line to the EOF character (Ctrl+D on Unix,
Ctrl+Z, Enter on Windows). This makes the program more robust when it comes
to handling input files that contain blank lines. We have stopped printing a
prompt for each number since it doesn’t make sense to have one for redirected
input. And we have also used a single try block with two exception handlers.
Notice that if an invalid integer is entered (either via the keyboard or due to
a “bad” line of data in a redirected input file), the int() conversion will raise a
ValueError exception and the flow of control will immediately switch to the relevant except block—this means that neither total nor count will be incremented
when invalid data is encountered, which is exactly what we want.
We could just as easily have used two separate exception-handling try blocks
while True:
line = input()
if line:
number = int(line)
except ValueError as err:
total += number
count += 1
except EOFError:
But we preferred to group the exceptions together at the end to keep the main
processing as uncluttered as possible.
Piece #8: Creating and Calling Functions
It is perfectly possible to write programs using the data types and control structures that we have covered in the preceding pieces. However, very often we
want to do essentially the same processing repeatedly, but with a small difference, such as a different starting value. Python provides a means of
Python’s “Beautiful Heart”
lating suites as functions which can be parameterized by the arguments they
are passed. Here is the general syntax for creating a function:
def functionName(arguments):
The arguments are optional and multiple arguments must be comma-separated.
Every Python function has a return value; this defaults to None unless we return
from the function using the syntax return value, in which case value is returned.
The return value can be just one value or a tuple of values. The return value
can be ignored by the caller, in which case it is simply thrown away.
Note that def is a statement that works in a similar way to the assignment
operator. When def is executed a function object is created and an object
reference with the specified name is created and set to refer to the function
object. Since functions are objects, they can be stored in collection data types
and passed as arguments to other functions, as we will see in later chapters.
One frequent need when writing interactive console applications is to obtain
an integer from the user. Here is a function that does just that:
def get_int(msg):
while True:
i = int(input(msg))
return i
except ValueError as err:
This function takes one argument, msg. Inside the while loop the user is prompted to enter an integer. If they enter something invalid a ValueError exception
will be raised, the error message will be printed, and the loop will repeat. Once
a valid integer is entered, it is returned to the caller. Here is how we would
call it:
age = get_int("enter your age: ")
In this example the single argument is mandatory because we have provided
no default value. In fact, Python supports a very sophisticated and flexible
syntax for function parameters that supports default argument values and
positional and keyword arguments. All of the syntax is covered in Chapter 4.
Although creating our own functions can be very satisfying, in many cases it
is not necessary. This is because Python has a lot of functions built in, and a
great many more functions in the modules in its standard library, so what we
want may well already be available.
➤ 173
Chapter 1. Rapid Introduction to Procedural Programming
A Python module is just a .py file that contains Python code, such as custom
function and class (custom data type) definitions, and sometimes variables. To
access the functionality in a module we must import it. For example:
import sys
To import a module we use the import statement followed by the name of the
.py file, but omitting the extension.★ Once a module has been imported, we can
access any functions, classes, or variables that it contains. For example:
The sys module provides the argv variable—a list whose first item is the name
under which the program was invoked and whose second and subsequent
items are the program’s command-line arguments. The two previously quoted
lines constitute the entire program. If the program is invoked
with the command line -v, it will print ['', '-v'] on the
console. (On Unix the first entry may be './'.)
Dot (.)
In general, the syntax for using a function from a module is moduleName.functionName(arguments). It makes use of the dot (“access attribute”) operator we
21 ➤
introduced in Piece #3. The standard library contains lots of modules, and we
will make use of many of them throughout the book. The standard modules
all have lowercase names, so some programmers use title-case names (e.g., MyModule) for their own modules to keep them distinct.
Let us look at just one example, the random module (in the standard library’s file), which provides many useful functions:
import random
x = random.randint(1, 6)
y = random.choice(["apple", "banana", "cherry", "durian"])
After these statements have been executed, x will contain an integer between
1 and 6 inclusive, and y will contain one of the strings from the list passed to
the random.choice() function.
(#!) line
12 ➤
It is conventional to put all the import statements at the beginning of .py files,
after the shebang line, and after the module’s documentation. (Documenting modules is covered in Chapter 5.) We recommend importing standard library modules first, then third-party library modules, and finally your own
The sys module, some other built-in modules, and modules implemented in C don’t necessarily
have corresponding .py files—but they are used in just the same way as those that do.
In the preceding section we learned enough Python to write real programs.
In this section we will study two complete programs that use only the Python
covered earlier. This is both to show what is possible, and to help consolidate
what has been learned so far.
In subsequent chapters we will increasingly cover more of Python’s language
and library, so that we will be able to write programs that are more concise and
more robust than those shown here—but first we must have the foundations
on which to build.
The first program we will review is quite short, although it has some subtle
aspects, including a list of lists. Here is what it does: Given a number on the
command line, the program outputs the same number onto the console using
“big” digits.
At sites where lots of users share a high-speed line printer, it used to be
common practice for each user’s print job to be preceded by a cover page that
showed their username and some other identifying details printed using this
kind of technique.
We will review the code in three parts: the import, the creation of the lists
holding the data the program uses, and the processing itself. But first, let’s
look at a sample run:
* *
* *
We have not shown the console prompt (or the leading ./ for Unix users); we
will take them for granted from now on.
import sys
Since we must read in an argument from the command line (the number
to output), we need to access the sys.argv list, so we begin by importing the
sys module.
We represent each number as a list of strings. For example, here is zero:
Chapter 1. Rapid Introduction to Procedural Programming
Zero = [" *** ",
" *
* ",
" *
* ",
" *** "]
One detail to note is that the Zero list of strings is spread over multiple lines.
Python statements normally occupy a single line, but they can span multiple
lines if they are a parenthesized expression, a list, set, or dictionary literal, a
function call argument list, or a multiline statement where every end-of-line
character except the last is escaped by preceding it with a backslash (\). In
all these cases any number of lines can be spanned and indentation does not
matter for the second and subsequent lines.
Each list representing a number has seven strings, all of uniform width,
although what this width is differs from number to number. The lists for the
other numbers follow the same pattern as for zero, although they are laid out
for compactness rather than for clarity:
One = [" * ", "** ", " * ", " * ", " * ", " * ", "***"]
Two = [" *** ", "*
*", "* * ", " * ", " *
", "*
# ...
Nine = [" ****", "*
*", "*
*", " ****", "
*", "
", "*****"]
*", "
The last piece of data we need is a list of all the lists of digits:
Digits = [Zero, One, Two, Three, Four, Five, Six, Seven, Eight, Nine]
We could have created the Digits lists directly, and avoided creating the extra
variables. For example:
Digits = [
[" *** ", " *
* ", "*
*", "*
*", "*
" *
* ", " *** "], # Zero
[" * ", "** ", " * ", " * ", " * ", " * ", "***"], # One
# ...
[" ****", "*
*", "*
*", " ****", "
*", "
*"] # Nine
We preferred to use a separate variable for each number both for ease of
understanding and because it looks neater using the variables.
We will quote the rest of the code in one go so that you can try to figure out how
it works before reading the explanation that follows.
set type
➤ 121
➤ 126
digits = sys.argv[1]
row = 0
while row < 7:
line = ""
column = 0
while column < len(digits):
number = int(digits[column])
digit = Digits[number]
line += digit[row] + " "
column += 1
row += 1
except IndexError:
print("usage: <number>")
except ValueError as err:
print(err, "in", digits)
The whole code is wrapped in an exception handler that can catch the two
things that can go wrong. We begin by retrieving the program’s command-line
argument. The sys.argv list is 0-based like all Python lists; the item at index
position 0 is the name the program was invoked as, so in a running program
this list always starts out with at least one item. If no argument was given we
will be trying to access the second item in a one-item list and this will cause
an IndexError exception to be raised. If this occurs, the flow of control is immediately switched to the corresponding exception-handling block, and there we
simply print the program’s usage. Execution then continues after the end of
the try block; but there is no more code, so the program simply terminates.
If no IndexError occurs, the digits string holds the command-line argument,
which we hope is a sequence of digit characters. (Remember from Piece #2 that
identifiers are case-sensitive, so digits and Digits are different.) Each big digit
is represented by seven strings, and to output the number correctly we must
output the top row of every digit, then the next row, and so on, until all seven
rows have been output. We use a while loop to iterate over each row. We could
just as easily have done this instead: for row in (0, 1, 2, 3, 4, 5, 6): and later
on we will see a much better way using the built-in range() function.
We use the line string to hold the row strings from all the digits involved. Then
we loop by column, that is, by each successive character in the command-line
argument. We retrieve each character with digits[column] and convert the
digit to an integer called number. If the conversion fails a ValueError exception
is raised and the flow of control immediately switches to the corresponding
exception handler. In this case we print an error message, and control resumes
after the try block. As noted earlier, since there is no more code at this point,
the program will simply terminate.
➤ 141
Chapter 1. Rapid Introduction to Procedural Programming
If the conversion succeeds, we use number as an index into the Digits list, from
which we extract the digit list of strings. We then add the row-th string from
this list to the line we are building up, and also append two spaces to give some
horizontal separation between each digit.
Each time the inner while loop finishes, we print the line that has been built
up. The key to understanding this program is where we append each digit’s
row string to the current row’s line. Try running the program to get a feel for
how it works. We will return to this program in the exercises to improve its
output slightly.
One frequently occurring need is the generation of test data. There is no single
generic program for doing this, since test data varies enormously. Python is
often used to produce test data because it is so easy to write and modify Python
programs. In this subsection we will create a program that generates a grid
of random integers; the user can specify how many rows and columns they
want and over what range the integers should span. We’ll start by looking at
a sample run:
rows: 4x
invalid literal for int() with base 10: '4x'
rows: 4
columns: 7
minimum (or Enter for 0): -100
maximum (or Enter for 1000):
The program works interactively, and at the beginning we made a typing error
when entering the number of rows. The program responded by printing an
error message and then asking us to enter the number of rows again. For the
maximum we just pressed Enter to accept the default.
We will review the code in four parts: the import, the definition of a get_int()
function (a more sophisticated version than the one shown in Piece #8), the
user interaction to get the values to use, and the processing itself.
import random
38 ➤
We need the random module to give us access to the random.randint() function.
def get_int(msg, minimum, default):
while True:
line = input(msg)
if not line and default is not None:
return default
i = int(line)
if i < minimum:
print("must be >=", minimum)
return i
except ValueError as err:
This function requires three arguments: a message string, a minimum value,
and a default value. If the user just presses Enter there are two possibilities. If
default is None, that is, no default value has been given, the flow of control will
drop through to the int() line. There the conversion will fail (since '' cannot
be converted to an integer), and a ValueError exception will be raised. But if
default is not None, then it is returned. Otherwise, the function will attempt
to convert the text the user entered into an integer, and if the conversion is
successful, it will then check that the integer is at least equal to the minimum
that has been specified.
So, the function will always return either default (if the user just pressed
Enter), or a valid integer that is greater than or equal to the specified minimum.
rows = get_int("rows: ", 1, None)
columns = get_int("columns: ", 1, None)
minimum = get_int("minimum (or Enter for 0): ", -1000000, 0)
default = 1000
if default < minimum:
default = 2 * minimum
maximum = get_int("maximum (or Enter for " + str(default) + "): ",
minimum, default)
Our get_int() function makes it easy to obtain the number of rows and
columns and the minimum random value that the user wants. For rows and
columns we give a default value of None, meaning no default, so the user must
enter an integer. In the case of the minimum, we supply a default value of 0,
and for the maximum we give a default value of 1 000, or twice the minimum
if the minimum is greater than or equal to 1 000.
As we noted in the previous example, function call argument lists can span
any number of lines, and indentation is irrelevant for their second and subsequent lines.
Chapter 1. Rapid Introduction to Procedural Programming
Once we know how many rows and columns the user requires and the minimum and maximum values of the random numbers they want, we are ready to
do the processing.
row = 0
while row < rows:
line = ""
column = 0
while column < columns:
i = random.randint(minimum, maximum)
s = str(i)
while len(s) < 10:
s = " " + s
line += s
column += 1
row += 1
To generate the grid we use three while loops, the outer one working by rows,
the middle one by columns, and the inner one by characters. In the middle
loop we obtain a random number in the specified range and then convert it to
a string. The inner while loop is used to pad the string with leading spaces so
that each number is represented by a string 10 characters wide. We use the
line string to accumulate the numbers for each row, and print the line after
each column’s numbers have been added. This completes our second example.
Python provides very sophisticated string formatting functionality, as well
as excellent support for for … in loops, so more realistic versions of both and would have used for … in loops, and would have used Python’s string formatting capabilities rather
than crudely padding with spaces. But we have limited ourselves to the eight
pieces of Python introduced in this chapter, and they are quite sufficient for
writing complete and useful programs. In each subsequent chapter we will
learn new Python features, so as we progress through the book the programs
we will see and be capable of writing will grow in sophistication.
In this chapter we learned how to edit and run Python programs and reviewed
a few small but complete programs. But most of the chapter’s pages were
devoted to the eight pieces of Python’s “beautiful heart”—enough of Python to
write real programs.
We began with two of Python’s most basic data types, int and str. Integer literals are written just as they are in most other programming languages. String
➤ 78
literals are written using single or double quotes; it doesn’t matter which as
long as the same kind of quote is used at both ends. We can convert between
strings and integers, for example, int("250") and str(125). If an integer conversion fails a ValueError exception is raised; whereas almost anything can be
converted to a string.
Strings are sequences, so those functions and operations that can be used with
sequences can be used with strings. For example, we can access a particular
character using the item access operator ([]), concatenate strings using +, and
append one string to another using +=. Since strings are immutable, behind
the scenes, appending creates a new string that is the concatenation of the
given strings, and rebinds the left-hand string object reference to the resultant
string. We can also iterate over a string character by character using a for … in
loop. And we can use the built-in len() function to report how many characters
are in a string.
For immutable objects like strings, integers, and tuples, we can write our code
as though an object reference is a variable, that is, as though an object reference is the object it refers to. We can also do this for mutable objects, although
any change made to a mutable object will affect all occurrences of the object
(i.e., all object references to the object); we will cover this issue in Chapter 3.
Python provides several built-in collection data types and has some others in its
standard library. We learned about the list and tuple types, and in particular
how to create tuples and lists from literals, for example, even = [2, 4, 6, 8]. Lists,
like everything else in Python, are objects, so we can call methods on them—for
example, even.append(10) will add an extra item to the list. Like strings, lists
and tuples are sequences, so we can iterate over them item by item using a
for … in loop, and find out how many items they have using len(). We can also
retrieve a particular item in a list or tuple using the item access operator ([]),
concatenate two lists or tuples using +, and append one to another using +=. If
we want to append a single item to a list we must either use list.append() or
use += with the item made into a single-item list—for example, even += [12].
Since lists are mutable, we can use [] to change individual items, for example,
even[1] = 16.
The fast is and is not identity operators can be used to check whether two object references refer to the same thing—this is particularly useful when checking against the unique built-in None object. All the usual comparison operators
are available (<, <=, ==, !=, >=, >), but they can be used only with compatible data
types, and then only if the operations are supported. The data types we have
seen so far—int, str, list, and tuple—all support the complete set of comparison operators. Comparing incompatible types, for example, comparing an int
with a str or list, will quite sensibly produce a TypeError exception.
Python supports the standard logical operators and, or, and not. Both and and
or are short-circuit operators that return the operand that determined their
Chapter 1. Rapid Introduction to Procedural Programming
result—and this may not be a Boolean (although it can be converted to a
Boolean); not always returns either True or False.
We can test for membership of sequence types, including strings, lists, and tuples, using the in and not in operators. Membership testing uses a slow linear
search on lists and tuples, and a potentially much faster hybrid algorithm for
strings, but performance is rarely an issue except for very long strings, lists,
and tuples. In Chapter 3 we will learn about Python’s associative array and
set collection data types, both of which provide very fast membership testing.
It is also possible to find out an object variable’s type (i.e., the type of object the
object reference refers to) using type()—but this function is normally used only
for debugging and testing.
Python provides several control structures, including conditional branching
with if … elif … else, conditional looping with while, looping over sequences
with for … in, and exception-handling with try … except blocks. Both while
and for … in loops can be prematurely terminated using a break statement,
and both can switch control to the beginning using continue.
The usual arithmetic operators are supported, including +, -, *, and /, although
Python is unusual in that / always produces a floating-point result even if both
its operands are integers. (The truncating division that many other languages
use is also available in Python as //.) Python also provides augmented assignment operators such as += and *=; these create new objects and rebind behind
the scenes if their left-hand operand is immutable. The arithmetic operators
are overloaded by the str and list types as we noted earlier.
Console I/O can be achieved using input() and print(); and using file redirection in the console, we can use these same built-in functions to read and
write files.
In addition to Python’s rich built-in functionality, its extensive standard
library is also available, with modules accessible once they have been imported
using the import statement. One commonly imported module is sys, which
holds the sys.argv list of command-line arguments. And when Python doesn’t
have the function we need we can easily create one that does what we want
using the def statement.
By making use of the functionality described in this chapter it is possible to
write short but useful Python programs. In the following chapter we will learn
more about Python’s data types, going into more depth for ints and strs and
introducing some entirely new data types. In Chapter 3 we will learn more
about tuples and lists, and also about some of Python’s other collection data
types. Then, in Chapter 4 we will cover Python’s control structures in much
more detail, and will learn how to create our own functions so that we can
package up functionality to avoid code duplication and promote code reuse.
The purpose of the exercises here, and throughout the book, is to encourage you
to experiment with Python, and to get hands-on experience to help you absorb
each chapter’s material. The examples and exercises cover both numeric and
text processing to appeal to as wide an audience as possible, and they are kept
fairly small so that the emphasis is on thinking and learning rather than just
typing code. Every exercise has a solution provided with the book’s examples.
1. One nice variation of the program is where instead of
printing *s, the relevant digit is printed instead. For example: 719428306
7 11
4 4
9999 4 4
9 444444
Two approaches can be taken. The easiest is to simply change the *s in
the lists. But this isn’t very versatile and is not the approach you should
take. Instead, change the processing code so that rather than adding each
digit’s row string to the line in one go, you add character by character, and
whenever a * is encountered you use the relevant digit.
This can be done by copying and changing about five lines.
It isn’t hard, but it is slightly subtle. A solution is provided as
2. IDLE can be used as a very powerful and flexible calculator, but sometimes it is useful to have a task-specific calculator. Create a program that
prompts the user to enter a number in a while loop, gradually building
up a list of the numbers entered. When the user has finished (by simply
pressing Enter), print out the numbers they entered, the count of numbers,
the sum of the numbers, the lowest and highest numbers entered, and the
mean of the numbers (sum / count). Here is a sample run:
enter a number or
enter a number or
enter a number or
enter a number or
enter a number or
enter a number or
enter a number or
Chapter 1. Rapid Introduction to Procedural Programming
numbers: [5, 4, 1, 8, 5, 2]
count = 6 sum = 25 lowest = 1 highest = 8 mean = 4.16666666667
It will take about four lines to initialize the necessary variables (an empty
list is simply []), and less than 15 lines for the while loop, including basic
error handling. Printing out at the end can be done in just a few lines, so
the whole program, including blank lines for the sake of clarity, should be
about 25 lines.
3. In some situations we need to generate test text—for example, to populate
a web site design before the real content is available, or to provide test
content when developing a report writer. To this end, write a program that
generates awful poems (the kind that would make a Vogon blush).
38 ➤
Create some lists of words, for example, articles (“the”, “a”, etc.), subjects
(“cat”, “dog”, “man”, “woman”), verbs (“sang”, “ran”, “jumped”), and adverbs
(“loudly”, “quietly”, “well”, “badly”). Then loop five times, and on each iteration use the random.choice() function to pick an article, subject, verb,
and adverb. Use random.randint() to choose between two sentence structures: article, subject, verb, and adverb, or just article, subject, and verb,
and print the sentence. Here is an example run:
another boy laughed badly
the woman jumped
a boy hoped
a horse jumped
another man laughed rudely
You will need to import the random module. The lists can be done in about
4–10 lines depending on how many words you put in them, and the loop
itself requires less than ten lines, so with some blank lines the whole
program can be done in about 20 lines of code. A solution is provided as
4. To make the awful poetry program more versatile, add some code to it so
that if the user enters a number on the command line (between 1 and 10
inclusive), the program will output that many lines. If no command-line
argument is given, default to printing five lines as before. You’ll need to
change the main loop (e.g., to a while loop). Keep in mind that Python’s
comparison operators can be chained, so there’s no need to use logical and
when checking that the argument is in range. The additional functionality
can be done by adding about ten lines of code. A solution is provided as
5. It would be nice to be able to calculate the median (middle value) as well
as the mean for the averages program in Exercise 2, but to do this we must
sort the list. In Python a list can easily be sorted using the list.sort()
method, but we haven’t covered that yet, so we won’t use it here. Extend the averages program with a block of code that sorts the list of
numbers—efficiency is of no concern, just use the easiest approach you
can think of. Once the list is sorted, the median is the middle value if the
list has an odd number of items, or the average of the two middle values
if the list has an even number of items. Calculate the median and output
that along with the other information.
This is rather tricky, especially for inexperienced programmers. If you
have some Python experience, you might still find it challenging, at least if
you keep to the constraint of using only the Python we have covered so far.
The sorting can be done in about a dozen lines and the median calculation
(where you can’t use the modulus operator, since it hasn’t been covered yet)
in four lines. A solution is provided in
● Identifiers and Keywords
● Integral Types
● Floating-Point Types
● Strings
Data Types
In this chapter we begin to take a much more detailed look at the Python language. We start with a discussion of the rules governing the names we give to
object references, and provide a list of Python’s keywords. Then we look at all
of Python’s most important data types—excluding collection data types, which
are covered in Chapter 3. The data types considered are all built-in, except for
one which comes from the standard library. The only difference between builtin data types and library data types is that in the latter case, we must first import the relevant module and we must qualify the data type’s name with the
name of the module it comes from—Chapter 5 covers importing in depth.
Identifiers and Keywords
16 ➤
When we create a data item we can either assign it to a variable, or insert it
into a collection. (As we noted in the preceding chapter, when we assign in
Python, what really happens is that we bind an object reference to refer to
the object in memory that holds the data.) The names we give to our object
references are called identifiers or just plain names.
A valid Python identifier is a nonempty sequence of characters of any length
that consists of a “start character” and zero or more “continuation characters”.
Such an identifier must obey a couple of rules and ought to follow certain conventions.
The first rule concerns the start and continuation characters. The start character can be anything that Unicode considers to be a letter, including the ASCII
letters (“a”, “b”, …, “z”, “A”, “B”, …, “Z”), the underscore (“_”), as well as the letters from most non-English languages. Each continuation character can be
any character that is permitted as a start character, or pretty well any nonwhitespace character, including any character that Unicode considers to be a
digit, such as (“0”, “1”, …, “9”), or the Catalan character “·”. Identifiers are case51
Chapter 2. Data Types
sensitive, so for example, TAXRATE, Taxrate, TaxRate, taxRate, and taxrate are five
different identifiers.
The precise set of characters that are permitted for the start and continuation
are described in the documentation (Python language reference, Lexical analysis, Identifiers and keywords section), and in PEP 3131 (Supporting Non-ASCII
The second rule is that no identifier can have the same name as one of Python’s
keywords, so we cannot use any of the names shown in Table 2.1.
Table 2.1 Python’s Keywords
We already met most of these keywords in the preceding chapter, although 11
of them—assert, class, del, finally, from, global, lambda, nonlocal, raise, with,
and yield—have yet to be discussed.
The first convention is: Don’t use the names of any of Python’s predefined identifiers for your own identifiers. So, avoid using NotImplemented and Ellipsis,
and the name of any of Python’s built-in data types (such as int, float, list,
str, and tuple), and any of Python’s built-in functions or exceptions. How can
we tell whether an identifier falls into one of these categories? Python has a
built-in function called dir() that returns a list of an object’s attributes. If it is
called with no arguments it returns the list of Python’s built-in attributes. For
>>> dir() # Python 3.1's list has an extra item, '__package__'
['__builtins__', '__doc__', '__name__']
The __builtins__ attribute is, in effect, a module that holds all of Python’s
built-in attributes. We can use it as an argument to the dir() function:
>>> dir(__builtins__)
['ArithmeticError', 'AssertionError', 'AttributeError',
'sum', 'super', 'tuple', 'type', 'vars', 'zip']
A “PEP” is a Python Enhancement Proposal. If someone wants to change or extend Python,
providing they get enough support from the Python community, they submit a PEP with the details
of their proposal so that it can be formally considered, and in some cases such as with PEP 3131,
accepted and implemented. All the PEPs are accessible from
Identifiers and Keywords
There are about 130 names in the list, so we have omitted most of them. Those
that begin with a capital letter are the names of Python’s built-in exceptions;
the rest are function and data type names.
The second convention concerns the use of underscores (_). Names that begin
and end with two underscores (such as __lt__) should not be used. Python
defines various special methods and variables that use such names (and in the
case of special methods, we can reimplement them, that is, make our own versions of them), but we should not introduce new names of this kind ourselves.
We will cover such names in Chapter 6. Names that begin with one or two leading underscores (and that don’t end with two underscores) are treated specially
in some contexts. We will show when to use names with a single leading underscore in Chapter 5, and when to use those with two leading underscores in
Chapter 6.
A single underscore on its own can be used as an identifier, and inside an
interactive interpreter or Python Shell, _ holds the result of the last expression
that was evaluated. In a normal running program no _ exists, unless we use it
explicitly in our code. Some programmers like to use _ in for … in loops when
they don’t care about the items being looped over. For example:
for _ in (0, 1, 2, 3, 4, 5):
38 ➤
Be aware, however, that those who write programs that are internationalized often use _ as the name of their translation function. They do this so
that instead of writing gettext.gettext("Translate me"), they can write
_("Translate me"). (For this code to work we must have first imported the gettext module so that we can access the module’s gettext() function.)
Let’s look at some valid identifiers in a snippet of code written by a Spanishspeaking programmer. The code assumes we have done import math and that
the variables radio and vieja_área have been created earlier in the program:
π = math.pi
ε = 0.0000001
nueva_área = π * radio * radio
if abs(nueva_área - vieja_área) < ε:
print("las áreas han convergido")
We’ve used the math module, set epsilon (ε) to be a very small floating-point
number, and used the abs() function to get the absolute value of the difference
between the areas—we cover all of these later in this chapter. What we are
concerned with here is that we are free to use accented characters and Greek
letters for identifiers. We could just as easily create identifiers using Arabic,
Chinese, Hebrew, Japanese, and Russian characters, or indeed characters from
any other language supported by the Unicode character set.
➤ 196
Chapter 2. Data Types
The easiest way to check whether something is a valid identifier is to try to
assign to it in an interactive Python interpreter or in IDLE’s Python Shell
window. Here are some examples:
>>> stretch-factor = 1
SyntaxError: can't assign to operator (...)
>>> 2miles = 2
SyntaxError: invalid syntax (...)
>>> str = 3 # Legal but BAD
>>> l'impôt31 = 4
SyntaxError: EOL while scanning single-quoted string (...)
>>> l_impôt31 = 5
When an invalid identifier is used it causes a SyntaxError exception to be raised.
In each case the part of the error message that appears in parentheses varies,
so we have replaced it with an ellipsis. The first assignment fails because
“-” is not a Unicode letter, digit, or underscore. The second one fails because
the start character is not a Unicode letter or underscore; only continuation
characters can be digits. No exception is raised if we create an identifier that
is valid—even if the identifier is the name of a built-in data type, exception,
or function—so the third assignment works, although it is ill-advised. The
fourth fails because a quote is not a Unicode letter, digit, or underscore. The
fifth is fine.
Integral Types
Python provides two built-in integral types, int and bool.★ Both integers and
Booleans are immutable, but thanks to Python’s augmented assignment operators this is rarely noticeable. When used in Boolean expressions, 0 and False
are False, and any other integer and True are True. When used in numerical
expressions True evaluates to 1 and False to 0. This means that we can write
some rather odd things—for example, we can increment an integer, i, using the
expression i += True. Naturally, the correct way to do this is i += 1.
The size of an integer is limited only by the machine’s memory, so integers
hundreds of digits long can easily be created and worked with—although they
will be slower to use than integers that can be represented natively by the
machine’s processor.
The standard library also provides the fractions.Fraction type (unlimited precision rationals)
which may be useful in some specialized mathematical and scientific contexts.
Dealing with
➤ 414
Integral Types
Table 2.2 Numeric Operators and Functions
x + y
Adds number x and number y
x - y
Subtracts y from x
x * y
Multiplies x by y
x / y
Divides x by y; always produces a float (or a complex if x or y
is complex)
x // y
Divides x by y; truncates any fractional part so always produces an int result; see also the round() function
x % y
Produces the modulus (remainder) of dividing x by y
x ** y
Raises x to the power of y; see also the pow() functions
Negates x; changes x’s sign if nonzero, does nothing if zero
Does nothing; is sometimes used to clarify code
Returns the absolute value of x
divmod(x, y)
Returns the quotient and remainder of dividing x by y as a
tuple of two ints
pow(x, y)
Raises x to the power of y; the same as the ** operator
pow(x, y, z)
A faster alternative to (x ** y) % z
round(x, n)
Returns x rounded to n integral digits if n is a negative int
or returns x rounded to n decimal places if n is a positive int;
the returned value has the same type as x; see the text
18 ➤
Table 2.3 Integer Conversion Functions
Returns the binary representation of int i as a string, e.g.,
bin(1980) == '0b11110111100'
Returns the hexadecimal representation of i as a string, e.g.,
hex(1980) == '0x7bc'
Converts object x to an integer; raises ValueError on
failure—or TypeError if x’s data type does not support integer
conversion. If x is a floating-point number it is truncated.
int(s, base)
Converts str s to an integer; raises ValueError on failure. If
the optional base argument is given it should be an integer
between 2 and 36 inclusive.
Returns the octal representation of i as a string, e.g.,
oct(1980) == '0o3674'
➤ 108
Chapter 2. Data Types
Integer literals are written using base 10 (decimal) by default, but other
number bases can be used when this is convenient:
>>> 14600926
>>> 0b110111101100101011011110
>>> 0o67545336
>>> 0xDECADE
# decimal
# binary
# octal
# hexadecimal
Binary numbers are written with a leading 0b, octal numbers with a leading
0o,★ and hexadecimal numbers with a leading 0x. Uppercase letters can also
be used.
All the usual mathematical functions and operators can be used with integers,
as Table 2.2 shows. Some of the functionality is provided by built-in functions
like abs()—for example, abs(i) returns the absolute value of integer i—and
other functionality is provided by int operators—for example, i + j returns the
sum of integers i and j.
We will mention just one of the functions from Table 2.2, since all the others are
sufficiently explained in the table itself. While for floats, the round() function
works in the expected way—for example, round(1.246, 2) produces 1.25—for
ints, using a positive rounding value has no effect and results in the same
number being returned, since there are no decimal digits to work on. But when
a negative rounding value is used a subtle and useful behavior is achieved—for
example, round(13579, -3) produces 14000, and round(34.8, -1) produces 30.0.
All the binary numeric operators (+, -, /, //, %, and **) have augmented assignment versions (+=, -=, /=, //=, %=, and **=) where x op= y is logically equivalent to
x = x op y in the normal case when reading x’s value has no side effects.
Objects can be created by assigning literals to variables, for example, x = 17, or
by calling the relevant data type as a function, for example, x = int(17). Some
objects (e.g., those of type decimal.Decimal) can be created only by using the
data type since they have no literal representation. When an object is created
using its data type there are three possible use cases.
The first use case is when a data type is called with no arguments. In this case
an object with a default value is created—for example, x = int() creates an
integer of value 0. All the built-in types can be called with no arguments.
The second use case is when the data type is called with a single argument. If
an argument of the same type is given, a new object which is a shallow copy of
Users of C-style languages note that a single leading 0 is not sufficient to specify an octal number;
0o (zero, letter o) must be used in Python.
Integral Types
the original object is created. (Shallow copying is covered in Chapter 3.) If an
argument of a different type is given, a conversion is attempted. This use is
shown for the int type in Table 2.3. If the argument is of a type that supports
conversions to the given type and the conversion fails, a ValueError exception
is raised; otherwise, the resultant object of the given type is returned. If the
argument’s data type does not support conversion to the given type a TypeError
exception is raised. The built-in float and str types both provide integer
conversions; it is also possible to provide integer and other conversions for our
own custom data types as we will see in Chapter 6.
➤ 146
➤ 252
The third use case is where two or more arguments are given—not all types
support this, and for those that do the argument types and their meanings
vary. For the int type two arguments are permitted where the first is a string
that represents an integer and the second is the number base of the string
representation. For example, int("A4", 16) creates an integer of value 164.
This use is shown in Table 2.3.
The bitwise operators are shown in Table 2.4. All the binary bitwise operators
(|, ^, &, <<, and >>) have augmented assignment versions (|=, ^=, &=, <<=, and
>>=) where i op= j is logically equivalent to i = i op j in the normal case when
reading i’s value has no side effects.
From Python 3.1, the int.bit_length() method is available. This returns
the number of bits required to represent the int it is called on. For example,
(2145).bit_length() returns 12. (The parentheses are required if a literal integer is used, but not if we use an integer variable.)
If many true/false flags need to be held, one possibility is to use a single integer,
and to test individual bits using the bitwise operators. The same thing can be
achieved less compactly, but more conveniently, using a list of Booleans.
Table 2.4 Integer Bitwise Operators
i | j
Bitwise OR of int i and int j; negative numbers are assumed to be
represented using 2’s complement
i ^ j
Bitwise XOR (exclusive or) of i and j
i & j
Bitwise AND of i and j
i << j
Shifts i left by j bits; like i * (2 ** j) without overflow checking
i >> j
Shifts i right by j bits; like i // (2 ** j) without overflow checking
Inverts i’s bits
Chapter 2. Data Types
There are two built-in Boolean objects: True and False. Like all other Python
data types (whether built-in, library, or custom), the bool data type can be
called as a function—with no arguments it returns False, with a bool argument
it returns a copy of the argument, and with any other argument it attempts
to convert the given object to a bool. All the built-in and standard library data
types can be converted to produce a Boolean value, and it is easy to provide
Boolean conversions for custom data types. Here are a couple of Boolean
assignments and a couple of Boolean expressions:
>>> t
>>> f
>>> t
>>> t
25 ➤
= True
= False
and f
and True
As we noted earlier, Python provides three logical operators: and, or, and not.
Both and and or use short-circuit logic and return the operand that determined
the result, whereas not always returns either True or False.
Programmers who have been using older versions of Python sometimes use
1 and 0 instead of True and False; this almost always works fine, but new code
should use the built-in Boolean objects when a Boolean value is required.
Floating-Point Types
Python provides three kinds of floating-point values: the built-in float and
complex types, and the decimal.Decimal type from the standard library. All three
are immutable. Type float holds double-precision floating-point numbers
whose range depends on the C (or C# or Java) compiler Python was built with;
they have limited precision and cannot reliably be compared for equality.
Numbers of type float are written with a decimal point, or using exponential
notation, for example, 0.0, 4., 5.7, -2.5, -2e9, 8.9e-4.
Computers natively represent floating-point numbers using base 2—this
means that some decimals can be represented exactly (such as 0.5), but others
only approximately (such as 0.1 and 0.2). Furthermore, the representation uses
a fixed number of bits, so there is a limit to the number of digits that can be
held. Here is a salutary example typed into IDLE:
>>> 0.0, 5.4, -2.5, 8.9e-4
(0.0, 5.4000000000000004, -2.5, 0.00088999999999999995)
Floating-Point Types
The inexactness is not a problem specific to Python—all programming languages have this problem with floating-point numbers.
Python 3.1 produces much more sensible-looking output:
>>> 0.0, 5.4, -2.5, 8.9e-4
(0.0, 5.4, -2.5, 0.00089)
When Python 3.1 outputs a floating-point number, in most cases it uses David
Gay’s algorithm. This outputs the fewest possible digits without losing any
accuracy. Although this produces nicer output, it doesn’t change the fact
that computers (no matter what computer language is used) effectively store
floating-point numbers as approximations.
If we need really high precision there are two approaches we can take. One
approach is to use ints—for example, working in terms of pennies or tenths of
a penny or similar—and scale the numbers when necessary. This requires us
to be quite careful, especially when dividing or taking percentages. The other
approach is to use Python’s decimal.Decimal numbers from the decimal module.
These perform calculations that are accurate to the level of precision we specify
(by default, to 28 decimal places) and can represent periodic numbers like 0.1
exactly; but processing is a lot slower than with floats. Because of their accuracy, decimal.Decimal numbers are suitable for financial calculations.
Mixed mode arithmetic is supported such that using an int and a float produces a float, and using a float and a complex produces a complex. Because decimal.Decimals are of fixed precision they can be used only with other decimal.
Decimals and with ints, in the latter case producing a decimal.Decimal result.
If an operation is attempted using incompatible types, a TypeError exception
is raised.
Floating-Point Numbers
All the numeric operators and functions in Table 2.2 (55 ➤) can be used with
floats, including the augmented assignment versions. The float data type can
be called as a function—with no arguments it returns 0.0, with a float argument it returns a copy of the argument, and with any other argument it attempts to convert the given object to a float. When used for conversions a string
argument can be given, either using simple decimal notation or using exponential notation. It is possible that NaN (“not a number”) or “infinity” may be
produced by a calculation involving floats—unfortunately the behavior is not
consistent across implementations and may differ depending on the system’s
underlying math library.
Here is a simple function for comparing floats for equality to the limit of the
machine’s accuracy:
Chapter 2. Data Types
Table 2.5 The Math Module’s Functions and Constants #1
Returns the arc cosine of x in radians
Returns the arc hyperbolic cosine of x in radians
Returns the arc sine of x in radians
Returns the arc hyperbolic sine of x in radians
Returns the arc tangent of x in radians
math.atan2(y, x)
Returns the arc tangent of y / x in radians
Returns the arc hyperbolic tangent of x in radians
Returns ⎡x ⎤ , i.e., the smallest integer greater than or
equal to x as an int; e.g., math.ceil(5.4) == 6
Returns x with y’s sign
Returns the cosine of x in radians
Returns the hyperbolic cosine of x in radians
Converts float r from radians to degrees
The constant e; approximately 2.718 281 828 459 045 1
Returns ex, i.e., math.e ** x
Returns | x |, i.e., the absolute value of x as a float
Returns x!
Returns ⎣x ⎦ , i.e., the largest integer less than or equal
to x as an int; e.g., math.floor(5.4) == 5
math.fmod(x, y)
Produces the modulus (remainder) of dividing x by y;
this produces better results than % for floats
Returns a 2-tuple with the mantissa (as a float) and
the exponent (as an int) so, x = m × 2 ; see math.ldexp()
Returns the sum of the values in iterable i as a float
math.hypot(x, y)
Returns √ x2 + y2
18 ➤
Returns True if float x is ± inf ( ± ∞)
Returns True if float x is nan (“not a number”)
math.ldexp(m, e)
Returns m × 2 ; effectively the inverse of math.frexp()
math.log(x, b)
Returns logb x; b is optional and defaults to math.e
Returns log10x
Returns loge(1 + x); accurate even when x is close to 0
Returns x’s fractional and whole parts as two floats
➤ 108
Floating-Point Types
Table 2.6 The Math Module’s Functions and Constants #2
The constant π; approximately 3.141 592 653 589 793 1
math.pow(x, y)
Returns xy as a float
Converts float d from degrees to radians
Returns the sine of x in radians
Returns the hyperbolic sine of x in radians
Returns √ x
Returns the tangent of x in radians
Returns the hyperbolic tangent of x in radians
Returns the whole part of x as an int; same as int(x)
def equal_float(a, b):
return abs(a - b) <= sys.float_info.epsilon
This requires us to import the sys module. The sys.float_info object has many
attributes; sys.float_info.epsilon is effectively the smallest difference that the
machine can distinguish between two floating-point numbers. On one of the
author’s 32-bit machines it is just over 0.000 000 000 000 000 2. (Epsilon is the
traditional name for this number.) Python floats normally provide reliable
accuracy for up to 17 significant digits.
If you type sys.float_info into IDLE, all its attributes will be displayed; these
include the minimum and maximum floating-point numbers the machine can
represent. And typing help(sys.float_info) will print some information about
the sys.float_info object.
Floating-point numbers can be converted to integers using the int() function which returns the whole part and throws away the fractional part, or
using round() which accounts for the fractional part, or using math.floor()
or math.ceil() which convert down to or up to the nearest integer. The
float.is_integer() method returns True if a floating-point number’s fractional part is 0, and a float’s fractional representation can be obtained using
the float.as_integer_ratio() method. For example, given x = 2.75, the call
x.as_integer_ratio() returns (11, 4). Integers can be converted to floatingpoint numbers using float().
Floating-point numbers can also be represented as strings in hexadecimal
format using the float.hex() method. Such strings can be converted back to
floating-point numbers using the float.fromhex() method. For example:
s = 14.25.hex()
# str s == '0x1.c800000000000p+3'
Chapter 2. Data Types
f = float.fromhex(s)
t = f.hex()
# float f == 14.25
# str t == '0x1.c800000000000p+3'
The exponent is indicated using p (“power”) rather than e since e is a valid
hexadecimal digit.
In addition to the built-in floating-point functionality, the math module provides
many more functions that operate on floats, as shown in Tables 2.5 and 2.6.
Here are some code snippets that show how to make use of the module’s functionality:
>>> import math
>>> math.pi * (5 ** 2) # Python 3.1 outputs: 78.53981633974483
>>> math.hypot(5, 12)
>>> math.modf(13.732) # Python 3.1 outputs: (0.7319999999999993, 13.0)
(0.73199999999999932, 13.0)
The math.hypot() function calculates the distance from the origin to the point
(x, y) and produces the same result as math.sqrt((x ** 2) + (y ** 2)).
The math module is very dependent on the underlying math library that Python
was compiled against. This means that some error conditions and boundary
cases may behave differently on different platforms.
Complex Numbers
The complex data type is an immutable type that holds a pair of floats, one
representing the real part and the other the imaginary part of a complex
number. Literal complex numbers are written with the real and imaginary
parts joined by a + or - sign, and with the imaginary part followed by a j.★ Here
are some examples: 3.5+2j, 0.5j, 4+0j, -1-3.7j. Notice that if the real part is 0,
we can omit it entirely.
The separate parts of a complex are available as attributes real and imag.
For example:
>>> z = -89.5+2.125j
>>> z.real, z.imag
(-89.5, 2.125)
Except for //, %, divmod(), and the three-argument pow(), all the numeric
operators and functions in Table 2.2 (55 ➤) can be used with complex numbers,
and so can the augmented assignment versions. In addition, complex numbers
Mathematicians use i to signify √ − 1, but Python follows the engineering tradition and uses j.
Floating-Point Types
have a method, conjugate(), which changes the sign of the imaginary part.
For example:
>>> z.conjugate()
>>> 3-4j.conjugate()
Notice that here we have called a method on a literal complex number. In general, Python allows us to call methods or access attributes on any literal, as long
as the literal’s data type provides the called method or the attribute—however,
this does not apply to special methods, since these always have corresponding
operators such as + that should be used instead. For example, 4j.real produces
0.0, 4j.imag produces 4.0, and 4j + 3+2j produces 3+6j.
The complex data type can be called as a function—with no arguments it
returns 0j, with a complex argument it returns a copy of the argument, and
with any other argument it attempts to convert the given object to a complex.
When used for conversions complex() accepts either a single string argument,
or one or two floats. If just one float is given, the imaginary part is taken to
be 0j.
The functions in the math module do not work with complex numbers. This is
a deliberate design decision that ensures that users of the math module get
exceptions rather than silently getting complex numbers in some situations.
Users of complex numbers can import the cmath module, which provides complex number versions of most of the trigonometric and logarithmic functions
that are in the math module, plus some complex number-specific functions such
as cmath.phase(), cmath.polar(), and cmath.rect(), and also the cmath.pi and
cmath.e constants which hold the same float values as their math module counterparts.
Decimal Numbers
In many applications the numerical inaccuracies that can occur when using
floats don’t matter, and in any case are far outweighed by the speed of calculation that floats offer. But in some cases we prefer the opposite trade-off, and
want complete accuracy, even at the cost of speed. The decimal module provides
immutable Decimal numbers that are as accurate as we specify. Calculations
involving Decimals are slower than those involving floats, but whether this is
noticeable will depend on the application.
To create a Decimal we must import the decimal module. For example:
>>> import decimal
>>> a = decimal.Decimal(9876)
Chapter 2. Data Types
>>> b = decimal.Decimal("54321.012345678987654321")
>>> a + b
Decimal numbers are created using the decimal.Decimal() function. This
function can take an integer or a string argument—but not a float, since floats
are held inexactly whereas decimals are represented exactly. If a string is
used it can use simple decimal notation or exponential notation. In addition
to providing accuracy, the exact representation of decimal.Decimals means that
they can be reliably compared for equality.
From Python 3.1 it is possible to convert floats to decimals using the decimal.Decimal.from_float() function. This function takes a float as argument
and returns the decimal.Decimal that is closest to the number the float approximates.
All the numeric operators and functions listed in Table 2.2 (55 ➤), including
the augmented assignment versions, can be used with decimal.Decimals, but
with a couple of caveats. If the ** operator has a decimal.Decimal left-hand
operand, its right-hand operand must be an integer. Similarly, if the pow()
function’s first argument is a decimal.Decimal, then its second and optional
third arguments must be integers.
The math and cmath modules are not suitable for use with decimal.Decimals,
but some of the functions provided by the math module are provided as decimal.Decimal methods. For example, to calculate ex where x is a float, we write
math.exp(x), but where x is a decimal.Decimal, we write x.exp(). From the discussion in Piece #3 (20 ➤), we can see that x.exp() is, in effect, syntactic sugar
for decimal.Decimal.exp(x).
The decimal.Decimal data type also provides ln() which calculates the natural
(base e) logarithm (just like math.log() with one argument), log10(), and sqrt(),
along with many other methods specific to the decimal.Decimal data type.
Numbers of type decimal.Decimal work within the scope of a context; the
context is a collection of settings that affect how decimal.Decimals behave. The
context specifies the precision that should be used (the default is 28 decimal
places), the rounding technique, and some other details.
In some situations the difference in accuracy between floats and decimal.
Decimals becomes obvious:
>>> 23 / 1.05
>>> print(23 / 1.05)
>>> print(decimal.Decimal(23) / decimal.Decimal("1.05"))
Floating-Point Types
>>> decimal.Decimal(23) / decimal.Decimal("1.05")
Although the division using decimal.Decimals is more accurate than the one
involving floats, in this case (on a 32-bit machine) the difference only shows
up in the fifteenth decimal place. In many situations this is insignificant—for
example, in this book, all the examples that need floating-point numbers use
One other point to note is that the last two of the preceding examples reveal
for the first time that printing an object involves some behind-the-scenes formatting. When we call print() on the result of decimal.Decimal(23) / decimal.Decimal("1.05") the bare number is printed—this output is in string form.
If we simply enter the expression we get a decimal.Decimal output—this output
is in representational form. All Python objects have two output forms. String
form is designed to be human-readable. Representational form is designed to
produce output that if fed to a Python interpreter would (when possible) reproduce the represented object. We will return to this topic in the next section
where we discuss strings, and again in Chapter 6 when we discuss providing
string and representational forms for our own custom data types.
The Library Reference’s decimal module documentation provides all the
details that are too obscure or beyond our scope to cover; it also provides more
examples, and a FAQ list.
Strings are represented by the immutable str data type which holds a sequence
of Unicode characters. The str data type can be called as a function to create
string objects—with no arguments it returns an empty string, with a nonstring argument it returns the string form of the argument, and with a string
argument it returns a copy of the string. The str() function can also be used
as a conversion function, in which case the first argument should be a string
or something convertable to a string, with up to two optional string arguments
being passed, one specifying the encoding to use and the other specifying how
to handle encoding errors.
Earlier we mentioned that string literals are created using quotes, and that we
are free to use single or double quotes providing we use the same at both ends.
In addition, we can use a triple quoted string—this is Python-speak for a string
that begins and ends with three quote characters (either three single quotes or
three double quotes). For example:
text = """A triple quoted string like this can include 'quotes' and
"quotes" without formality. We can also escape newlines \
so this particular string is actually only two lines long."""
➤ 91
Chapter 2. Data Types
Table 2.7 Python’s String Escapes
Escape (i.e., ignore) the newline
Backslash (\)
Single quote (’)
Double quote (")
ASCII bell (BEL)
ASCII backspace (BS)
ASCII formfeed (FF)
ASCII linefeed (LF)
Unicode character with the given name
Character with the given octal value
ASCII carriage return (CR)
Unicode character with the given 16-bit hexadecimal value
Unicode character with the given 32-bit hexadecimal value
ASCII vertical tab (VT)
Character with the given 8-bit hexadecimal value
If we want to use quotes inside a normal quoted string we can do so without
formality if they are different from the delimiting quotes; otherwise, we must
escape them:
a = "Single 'quotes' are fine; \"doubles\" must be escaped."
b = 'Single \'quotes\' must be escaped; "doubles" are fine.'
Python uses newline as its statement terminator, except inside parentheses
(()), square brackets ([]), braces ({}), or triple quoted strings. Newlines can be
used without formality in triple quoted strings, and we can include newlines
in any string literal using the \n escape sequence. All of Python’s escape sequences are shown in Table 2.7. In some situations—for example, when writing
regular expressions—we need to create strings with lots of literal backslashes.
(Regular expressions are the subject of Chapter 13.) This can be inconvenient
since each one must be escaped:
import re
phone1 = re.compile("^((?:[(]\\d+[)])?\\s*\\d+(?:-\\d+)?)$")
The solution is to use raw strings. These are quoted or triple quoted strings
whose first quote is preceded by the letter r. Inside such strings all characters
are taken to be literals, so no escaping is necessary. Here is the phone regular
expression using a raw string:
phone2 = re.compile(r"^((?:[(]\d+[)])?\s*\d+(?:-\d+)?)$")
If we want to write a long string literal spread over two or more lines but without using a triple quoted string there are a couple of approaches we can take:
t = "This is not the best way to join two long strings " + \
"together since it relies on ugly newline escaping"
s = ("This is the nice way to join two long strings "
"together; it relies on string literal concatenation.")
Notice that in the second case we must use parentheses to create a single
expression—without them, s would be assigned only to the first string, and
the second string would cause an IndentationError exception to be raised. The
Python documentation’s “Idioms and Anti-Idioms” HOWTO document recommends always using parentheses to spread statements of any kind over multiple lines rather than escaping newlines; a recommendation we endeavor to
Since .py files default to using the UTF-8 Unicode encoding, we can write any
Unicode characters in our string literals without formality. We can also put
any Unicode characters inside strings using hexadecimal escape sequences or
using Unicode names. For example:
>>> euros = " \N{euro sign} \u20AC \U000020AC"
>>> print(euros)
In this case we could not use a hexadecimal escape because they are limited to
two digits, so they cannot exceed 0xFF. Note that Unicode character names are
not case-sensitive, and spaces inside them are optional.
If we want to know the Unicode code point (the integer assigned to the character in the Unicode encoding) for a particular character in a string, we can use
the built-in ord() function. For example:
➤ 91
>>> ord(euros[0])
>>> hex(ord(euros[0]))
Similarly, we can convert any integer that represents a valid code point into
the corresponding Unicode character using the built-in chr() function:
Chapter 2. Data Types
>>> s = "anarchists are " + chr(8734) + chr(0x23B7)
>>> s
'anarchists are ∞√'
>>> ascii(s)
"'anarchists are \u221e\u23b7'"
If we enter s on its own in IDLE, it is output in its string form, which for strings
means the characters are output enclosed in quotes. If we want only ASCII
characters, we can use the built-in ascii() function which returns the representational form of its argument using 7-bit ASCII characters where possible, and
using the shortest form of \xhh, \uhhhh, or \Uhhhhhhhh escape otherwise. We will
see how to achieve precise control of string output later in this chapter.
➤ 78
Comparing Strings
Strings support the usual comparison operators <, <=, ==, !=, >, and >=. These
operators compare strings byte by byte in memory. Unfortunately, two problems arise when performing comparisons, such as when sorting lists of
strings. Both problems afflict every programming language that uses Unicode
strings—neither is specific to Python.
The first problem is that some Unicode characters can be represented by two
or more different byte sequences. For example, the character Å (Unicode code
point 0x00C5) can be represented in UTF-8 encoded bytes in three different
ways: [0xE2, 0x84, 0xAB], [0xC3, 0x85], and [0x41, 0xCC, 0x8A]. Fortunately, we
can solve this problem. If we import the unicodedata module and call unicodedata.normalize() with "NFKC" as the first argument (this is a normalization
method—three others are also available, "NFC", "NFD", and "NFKD"), and a string
containing the Å character using any of its valid byte sequences, the function
will return a string that when represented as UTF-8 encoded bytes will always
be the byte sequence [0xC3, 0x85].
The second problem is that the sorting of some characters is language-specific.
One example is that in Swedish ä is sorted after z, whereas in German, ä is sorted as if though were spelled ae. Another example is that although in English
we sort ø as if it were o, in Danish and Norwegian it is sorted after z. There
are lots of problems along these lines, and they can be complicated by the fact
that sometimes the same application is used by people of different nationalities
(who therefore expect different sorting orders), and sometimes strings are in a
mixture of languages (e.g., some Spanish, others English), and some characters
(such as arrows, dingbats, and mathematical symbols) don’t really have meaningful sort positions.
As a matter of policy—to prevent subtle mistakes—Python does not make
guesses. In the case of string comparisons, it compares using the strings’ inmemory byte representation. This gives a sort order based on Unicode code
➤ 91
points which gives ASCII sorting for English. Lower- or uppercasing all the
strings compared produces a more natural English language ordering. Normalizing is unlikely to be needed unless the strings are from external sources like
files or network sockets, but even in these cases it probably shouldn’t be done
unless there is evidence that it is needed. We can of course customize Python’s
sort methods as we will see in Chapter 3. The whole issue of sorting Unicode
strings is explained in detail in the Unicode Collation Algorithm document
Slicing and Striding Strings
Piece #3
18 ➤
We know from Piece #3 that individual items in a sequence, and therefore individual characters in a string, can be extracted using the item access operator
([]). In fact, this operator is much more versatile and can be used to extract not
just one item or character, but an entire slice (subsequence) of items or characters, in which context it is referred to as the slice operator.
First we will begin by looking at extracting individual characters. Index
positions into a string begin at 0 and go up to the length of the string minus
1. But it is also possible to use negative index positions—these count from the
last character back toward the first. Given the assignment s = "Light ray",
Figure 2.1 shows all the valid index positions for string s.
g h
a y
Figure 2.1 String index positions
Negative indexes are surprisingly useful, especially -1 which always gives us
the last character in a string. Accessing an out-of-range index (or any index in
an empty string) will cause an IndexError exception to be raised.
The slice operator has three syntaxes:
The seq can be any sequence, such as a list, string, or tuple. The start, end, and
step values must all be integers (or variables holding integers). We have used
the first syntax already: It extracts the start-th item from the sequence. The
second syntax extracts a slice from and including the start-th item, up to and
excluding the end-th item. We’ll discuss the third syntax shortly.
Chapter 2. Data Types
offers a better solution. The method takes a sequence as an argument (e.g., a
list or tuple of strings), and joins them together into a single string with the
string the method was called on between each one. For example:
>>> treatises = ["Arithmetica", "Conics", "Elements"]
>>> " ".join(treatises)
'Arithmetica Conics Elements'
>>> "-<>-".join(treatises)
>>> "".join(treatises)
The first example is perhaps the most common, joining with a single character,
in this case a space. The third example is pure concatenation thanks to the
empty string which means that the sequence of strings are joined with nothing
in between.
The str.join() method can also be used with the built-in reversed() function,
to reverse a string, for example, "".join(reversed(s)), although the same result
can be achieved more concisely by striding, for example, s[::-1].
The * operator provides string replication:
>>> s = "=" * 5
>>> print(s)
>>> s *= 10
>>> print(s)
As the example shows, we can also use the augmented assignment version of
the replication operator.★
When applied to strings, the in membership operator returns True if its lefthand string argument is a substring of, or equal to, its right-hand string argument.
In cases where we want to find the position of one string inside another, we
have two methods to choose from. One is the str.index() method; this returns
the index position of the substring, or raises a ValueError exception on failure.
The other is the str.find() method; this returns the index position of the substring, or -1 on failure. Both methods take the string to find as their first argument, and can accept a couple of optional arguments. The second argument
is the start position in the string being searched, and the third argument is the
end position in the string being searched.
Strings also support the % operator for formatting. This operator is deprecated and provided only
to ease conversion from Python 2 to Python 3. It is not used in any of the book’s examples.
Table 2.8 String Methods #1
Returns a copy of str s with the first letter capitalized;
see also the str.title() method
Returns a copy of s centered in a string of length width
padded with spaces or optionally with char (a string of
length 1); see str.ljust(), str.rjust(), and str.format(),
start, end)
Returns the number of occurrences of str t in str s (or in
the start:end slice of s)
Returns a bytes object that represents the string using
the default encoding or using the specified encoding and
handling errors according to the optional err argument
start, end)
Returns True if s (or the start:end slice of s) ends with str
x or with any of the strings in tuple x; otherwise, returns
False. See also str.startswith().
Returns a copy of s with tabs replaced with spaces in
multiples of 8 or of size if specified
➤ 91
start, end)
Returns the leftmost position of t in s (or in the start:end
slice of s) or -1 if not found. Use str.rfind() to find the
rightmost position. See also str.index().
Returns a copy of s formatted according to the given
arguments. This method and its arguments are covered
in the next subsection.
Returns the leftmost position of t in s (or in the
start:end slice of s) or raises ValueError if not found. Use
str.rindex() to search from the right. See str.find().
start, end)
Returns True if s is nonempty and every character in s
is alphanumeric
Returns True if s is nonempty and every character in s
is alphabetic
Returns True if s is nonempty and every character in s is
a Unicode base 10 digit
Returns True if s is nonempty and every character in s is
an ASCII digit
Returns True if s is nonempty and is a valid identifier
Returns True if s has at least one lowercaseable character and all its lowercaseable characters are lowercase;
see also str.isupper()
51 ➤
➤ 293
➤ 78
Chapter 2. Data Types
Table 2.9 String Methods #2
Returns True if s is nonempty and every character in s is
a numeric Unicode character such as a digit or fraction
Returns True if s is empty or if every character in s is considered to be printable, including space, but not newline
Returns True if s is nonempty and every character in s is
a whitespace character
Returns True if s is a nonempty title-cased string; see
also str.title()
Returns True if str s has at least one uppercaseable character and all its uppercaseable characters are uppercase;
see also str.islower()
Returns the concatenation of every item in the sequence
seq, with str s (which may be empty) between each one
Returns a copy of s left-aligned in a string of length width
padded with spaces or optionally with char (a string of
length 1). Use str.rjust() to right-align and
to center. See also str.format().
Returns a lowercased copy of s; see also str.upper()
Companion of str.translate(); see text for details
Returns a tuple of three strings—the part of str s before
the leftmost str t, t, and the part of s after t; or if t isn’t in
s returns s and two empty strings. Use str.rpartition()
to partition on the rightmost occurrence of t.
u, n)
Returns a copy of s with every (or a maximum of n if
given) occurrences of str t replaced with str u
s.split(t, n)
Returns a list of strings splitting at most n times on str t;
if n isn’t given, splits as many times as possible; if t isn’t
given, splits on whitespace. Use str.rsplit() to split from
the right—this makes a difference only if n is given and is
less than the maximum number of splits possible.
Returns the list of lines produced by splitting s on line
terminators, stripping the terminators unless f is True
x, start,
Returns True if s (or the start:end slice of s) starts with
str x or with any of the strings in tuple x; otherwise,
returns False. See also str.endswith().
Table 2.10 String Methods #3
Returns a copy of s with leading and trailing whitespace
(or the characters in str chars) removed; str.lstrip() strips
only at the start, and str.rstrip() strips only at the end
Returns a copy of s with uppercase characters lowercased
and lowercase characters uppercased; see also str.lower()
and str.upper()
Returns a copy of s where the first letter of each word
is uppercased and all other letters are lowercased; see
Companion of str.maketrans(); see text for details
Returns an uppercased copy of s; see also str.lower()
Returns a copy of s, which if shorter than w is padded with
leading zeros to make it w characters long
Which search method we use is purely a matter of taste and circumstance,
although if we are looking for multiple index positions, using the str.index()
method often produces cleaner code, as the following two equivalent functions
def extract_from_tag(tag, line):
def extract_from_tag(tag, line):
opener = "<" + tag + ">"
opener = "<" + tag + ">"
closer = "</" + tag + ">"
closer = "</" + tag + ">"
i = line.find(opener)
i = line.index(opener)
if i != -1:
start = i + len(opener)
start = i + len(opener)
j = line.index(closer, start)
j = line.find(closer, start)
return line[start:j]
if j != -1:
except ValueError:
return line[start:j]
return None
return None
Both versions of the extract_from_tag() function have exactly the same behavior. For example, extract_from_tag("red", "what a <red>rose</red> this is")
returns the string “rose”. The exception-handling version on the left separates
out the code that does what we want from the code that handles errors, and the
error return value version on the right intersperses what we want with error
The methods str.count(), str.endswith(), str.find(), str.rfind(), str.index(),
str.rindex(), and str.startswith() all accept up to two optional arguments: a
start position and an end position. Here are a couple of equivalences to put
this in context, assuming that s is a string:
Chapter 2. Data Types
s.count("m", 6) == s[6:].count("m")
s.count("m", 5, -3) == s[5:-3].count("m")
As we can see, the string methods that accept start and end indexes operate on
the slice of the string specified by those indexes.
Now we will look at another equivalence, this time to help clarify the behavior
of str.partition()—although we’ll actually use a str.rpartition() example:
result = s.rpartition("/")
i = s.rfind("/")
if i == -1:
result = "", "", s
result = s[:i], s[i], s[i + 1:]
The left- and right-hand code snippets are not quite equivalent because the
one on the right also creates a new variable, i. Notice that we can assign tuples
without formality, and that in both cases we looked for the rightmost occurrence of /. If s is the string "/usr/local/bin/firefox", both snippets produce the
same result: ('/usr/local/bin', '/', 'firefox').
We can use str.endswith() (and str.startswith()) with a single string argument, for example, s.startswith("From:"), or with a tuple of strings. Here is a
statement that uses both str.endswith() and str.lower() to print a filename if
it is a JPEG file:
if filename.lower().endswith((".jpg", ".jpeg")):
print(filename, "is a JPEG image")
The is*() methods such as isalpha() and isspace() return True if the string
they are called on has at least one character, and every character in the string
meets the criterion. For example:
>>> "917.5".isdigit(), "".isdigit(), "-2".isdigit(), "203".isdigit()
(False, False, False, True)
The is*() methods work on the basis of Unicode character classifications, so
for example, calling str.isdigit() on the strings "\N{circled digit two}03" and
"➁03" returns True for both of them. For this reason we cannot assume that a
string can be converted to an integer when isdigit() returns True.
When we receive strings from external sources (other programs, files, network
connections, and especially interactive users), the strings may have unwanted
leading and trailing whitespace. We can strip whitespace from the left using
str.lstrip(), from the right using str.rstrip(), or from both ends using
str.strip(). We can also give a string as an argument to the strip methods, in
which case every occurrence of every character given will be stripped from the
appropriate end or ends. For example:
>>> s = "\t no parking "
>>> s.lstrip(), s.rstrip(), s.strip()
('no parking ', '\t no parking', 'no parking')
>>> "<[unbracketed]>".strip("[](){}<>")
We can also replace strings within strings using the str.replace() method.
This method takes two string arguments, and returns a copy of the string it is
called on with every occurrence of the first string replaced with the second. If
the second argument is an empty string the effect is to delete every occurrence
of the first string. We will see examples of str.replace() and some other string
methods in the example in the Examples section toward the end of
the chapter.
One frequent requirement is to split a string into a list of strings. For example, we might have a text file of data with one record per line and each record’s
fields separated by asterisks. This can be done using the str.split() method
and passing in the string to split on as its first argument, and optionally the
maximum number of splits to make as the second argument. If we don’t specify the second argument, as many splits are made as possible. Here is an example:
>>> record = "Leo Tolstoy*1828-8-28*1910-11-20"
>>> fields = record.split("*")
>>> fields
['Leo Tolstoy', '1828-8-28', '1910-11-20']
Now we can use str.split() again on the date of birth and date of death to
calculate how long he lived (give or take a year):
>>> born = fields[1].split("-")
>>> born
['1828', '8', '28']
>>> died = fields[2].split("-")
>>> print("lived about", int(died[0]) - int(born[0]), "years")
lived about 82 years
We had to use int() to convert the years from strings to integers, but other than
that the snippet is straightforward. We could have gotten the years directly
from the fields list, for example, year_born = int(fields[1].split("-")[0]).
The two methods that we did not summarize in Tables 2.8, 2.9, and 2.10 are
str.maketrans() and str.translate(). The str.maketrans() method is used to
create a translation table which maps characters to characters. It accepts one,
two, or three arguments, but we will show only the simplest (two argument)
call where the first argument is a string containing characters to translate from
and the second argument is a string containing the characters to translate to.
➤ 97
Chapter 2. Data Types
Both arguments must be the same length. The str.translate() method takes
a translation table as an argument and returns a copy of its string with the
characters translated according to the translation table. Here is how we could
translate strings that might contain Bengali digits to English digits:
table = "".maketrans("\N{bengali digit zero}"
"\N{bengali digit one}\N{bengali digit two}"
"\N{bengali digit three}\N{bengali digit four}"
"\N{bengali digit five}\N{bengali digit six}"
"\N{bengali digit seven}\N{bengali digit eight}"
"\N{bengali digit nine}", "0123456789")
# prints: 20749
print("\N{bengali digit two}07\N{bengali digit four}"
"\N{bengali digit nine}".translate(table))
# prints: 20749
Notice that we have taken advantage of Python’s string literal concatenation
inside the str.maketrans() call and inside the second print() call to spread
strings over multiple lines without having to escape newlines or use explicit
We called str.maketrans() on an empty string because it doesn’t matter what
string it is called on; it simply processes its arguments and returns a translation table. The str.maketrans() and str.translate() methods can also be used
to delete characters by passing a string containing the unwanted characters as
the third argument to str.maketrans(). If more sophisticated character translations are required, we could create a custom codec—see the codecs module
documentation for more about this.
Python has a few other library modules that provide string-related functionality. We’ve already briefly mentioned the unicodedata module, and we’ll show
it in use in the next subsection. Other modules worth looking up are difflib
which can be used to show differences between files or between strings, the io
module’s io.StringIO class which allows us to read from or write to strings as
though they were files, and the textwrap module which provides facilities for
wrapping and filling strings. There is also a string module that has a few useful constants such as ascii_letters and ascii_lowercase. We will see examples
of some of these modules in use in Chapter 5. In addition, Python provides excellent support for regular expressions in the re module—Chapter 13 is dedicated to this topic.
String Formatting with the str.format() Method
The str.format() method provides a very flexible and powerful way of creating
strings. Using str.format() is easy for simple cases, but for complex formatting
we need to learn the formatting syntax the method requires.
The str.format() method returns a new string with the replacement fields in
its string replaced with its arguments suitably formatted. For example:
>>> "The novel '{0}' was published in {1}".format("Hard Times", 1854)
"The novel 'Hard Times' was published in 1854"
Each replacement field is identified by a field name in braces. If the field
name is a simple integer, it is taken to be the index position of one of the
arguments passed to str.format(). So in this case, the field whose name was 0
was replaced by the first argument, and the one with name 1 was replaced by
the second argument.
If we need to include braces inside format strings, we can do so by doubling
them up. Here is an example:
>>> "{{{0}}} {1} ;-}}".format("I'm in braces", "I'm not")
"{I'm in braces} I'm not ;-}"
If we try to concatenate a string and a number, Python will quite rightly raise
a TypeError. But we can easily achieve what we want using str.format():
>>> "{0}{1}".format("The amount due is $", 200)
'The amount due is $200'
We can also concatenate strings using str.format() (although the str.join()
method is best for this):
>>> x = "three"
>>> s ="{0} {1} {2}"
>>> s = s.format("The", x, "tops")
>>> s
'The three tops'
Here we have used a couple of string variables, but in most of this section
we’ll use string literals for str.format() examples, simply for the sake of
convenience—just keep in mind that any example that uses a string literal
could use a string variable in exactly the same way.
The replacement field can have any of the following general syntaxes:
One other point to note is that replacement fields can contain replacement
fields. Nested replacement fields cannot have any formatting; their purpose is
to allow for computed formatting specifications. We will see an example of this
Chapter 2. Data Types
when we take a detailed look at format specifications. We will now study each
part of the replacement field in turn, starting with field names.
Field Names
A field name can be either an integer corresponding to one of the str.format()
method’s arguments, or the name of one of the method’s keyword arguments.
We discuss keyword arguments in Chapter 4, but they are not difficult, so we
will provide a couple of examples here for completeness:
>>> "{who} turned {age} this year".format(who="She", age=88)
'She turned 88 this year'
>>> "The {who} was {0} last week".format(12, who="boy")
'The boy was 12 last week'
The first example uses two keyword arguments, who and age, and the second
example uses one positional argument (the only kind we have used up to
now) and one keyword argument. Notice that in an argument list, keyword
arguments always come after positional arguments; and of course we can make
use of any arguments in any order inside the format string.
Field names may refer to collection data types—for example, lists. In such
cases we can include an index (not a slice!) to identify a particular item:
>>> stock = ["paper", "envelopes", "notepads", "pens", "paper clips"]
>>> "We have {0[1]} and {0[2]} in stock".format(stock)
'We have envelopes and notepads in stock'
The 0 refers to the positional argument, so {0[1]} is the stock list argument’s
second item, and {0[2]} is the stock list argument’s third item.
Later on we will learn about Python dictionaries. These store key–value items,
and since they can be used with str.format(), we’ll just show a quick example
here. Don’t worry if it doesn’t make sense; it will once you’ve read Chapter 3.
>>> d = dict(animal="elephant", weight=12000)
>>> "The {0[animal]} weighs {0[weight]}kg".format(d)
'The elephant weighs 12000kg'
Just as we access list and tuple items using an integer position index, we access
dictionary items using a key.
We can also access named attributes. Assuming we have imported the math and
sys modules, we can do this:
>>> "math.pi=={0.pi} sys.maxunicode=={1.maxunicode}".format(math, sys)
'math.pi==3.14159265359 sys.maxunicode==65535'
➤ 126
So in summary, the field name syntax allows us to refer to positional and keyword arguments that are passed to the str.format() method. If the arguments
are collection data types like lists or dictionaries, or have attributes, we can access the part we want using [] or . notation. This is illustrated in Figure 2.5.
positional argument index
keyword argument name
Figure 2.5 Annotated format specifier field name examples
From Python 3.1 it is possible to omit field names, in which case Python will in
effect put them in for us, using numbers starting from 0. For example:
>>> "{} {} {}".format("Python", "can", "count")
'Python can count'
If we are using Python 3.0, the format string used here would have to be "{0}
{1} {2}". Using this technique is convenient for formatting one or two items,
but the approach we will look at next is more convenient when several items
are involved, and works just as well with Python 3.0.
Before finishing our discussion of string format field names, it is worth mentioning a rather different way to get values into a format string. This involves
using an advanced technique, but one useful to learn as soon as possible, since
it is so convenient.
The local variables that are currently in scope are available from the built-in
locals() function. This function returns a dictionary whose keys are local
variable names and whose values are references to the variables’ values. Now
we can use mapping unpacking to feed this dictionary into the str.format()
method. The mapping unpacking operator is ** and it can be applied to a
mapping (such as a dictionary) to produce a key–value list suitable for passing
to a function. For example:
>>> element = "Silver"
>>> number = 47
>>> "Element {number} is {element}".format(**locals())
'Element 47 is Silver'
➤ 179
Chapter 2. Data Types
The syntax may seem weird enough to make a Perl programmer feel at home,
but don’t worry—it is explained in Chapter 4. All that matters for now is that
we can use variable names in format strings and leave Python to fill in their
values simply by unpacking the dictionary returned by locals()—or some
other dictionary—into the str.format() method. For example, we could rewrite
the “elephant” example we saw earlier to have a much nicer format string with
simpler field names.
➤ 177
>>> "The {animal} weighs {weight}kg".format(**d)
'The elephant weighs 12000kg'
Unpacking a dictionary into the str.format() method allows us to use the
dictionary’s keys as field names. This makes string formats much easier to
understand, and also easier to maintain, since they are not dependent on the
order of the arguments. Note, however, that if we want to pass more than one
argument to str.format(), only the last one can use mapping unpacking.
63 ➤
When we discussed decimal.Decimal numbers we noticed that such numbers
are output in one of two ways. For example:
>>> decimal.Decimal("3.4084")
>>> print(decimal.Decimal("3.4084"))
The first way that the decimal.Decimal is shown is in its representational form.
The purpose of this form is to provide a string which if interpreted by Python
would re-create the object it represents. Python programs can evaluate snippets of Python code or entire programs, so this facility can be useful in some
situations. Not all objects can provide a reproducing representation, in which
case they provide a string enclosed in angle brackets. For example, the representational form of the sys module is the string "<module 'sys' (built-in)>".
The second way that decimal.Decimal is shown is in its string form. This form is
aimed at human readers, so the concern is to show something that makes sense
to people. If a data type doesn’t have a string form and a string is required,
Python will use the representational form.
Python’s built-in data types know about str.format(), and when passed as an
argument to this method they return a suitable string to display themselves.
It is also straightforward to add str.format() support to custom data types as
we will see in Chapter 6. In addition, it is possible to override the data type’s
normal behavior and force it to provide either its string or its representational
form. This is done by adding a conversion specifier to the field. Currently there
are three such specifiers: s to force string form, r to force representational form,
➤ 344
and a to force representational form but only using ASCII characters. Here is
an example:
>>> "{0} {0!s} {0!r} {0!a}".format(decimal.Decimal("93.4"))
"93.4 93.4 Decimal('93.4') Decimal('93.4')"
In this case, decimal.Decimal’s string form produces the same string as the
string it provides for str.format() which is what commonly happens. Also, in
this particular example, there is no difference between the representational
and ASCII representational forms since both use only ASCII characters.
Here is another example, this time concerning a string that contains the title of a movie, "
", held in the variable movie. If we print the
string using "{0}".format(movie) the string will be output unchanged, but
if we want to avoid non-ASCII characters we can use either ascii(movie) or
"{0!a}".format(movie), both of which will produce the string '\u7ffb\u8a33
So far we have seen how to put the values of variables into a format string, and
how to force string or representational forms to be used. Now we are ready to
consider the formatting of the values themselves.
Format Specifications
The default formatting of integers, floating-point numbers, and strings is often
perfectly satisfactory. But if we want to exercise fine control, we can easily do
so using format specifications. We will deal separately with formatting strings,
integers, and floating-point numbers, to make learning the details easier. The
the general syntax that covers all of them is shown in Figure 2.6.
For strings, the things that we can control are the fill character, the alignment
within the field, and the minimum and maximum field widths.
A string format specification is introduced with a colon (:) and this is followed
by an optional pair of characters—a fill character (which may not be }) and an
alignment character (< for left align, ^ for center, > for right align). Then comes
an optional minimum width integer, and if we want to specify a maximum
width, this comes last as a period followed by an integer.
Note that if we specify a fill character we must also specify an alignment. We
omit the sign and type parts of the format specification because they have no
effect on strings. It is harmless (but pointless) to have a colon without any of
the optional elements.
Let’s see some examples:
>>> s = "The sword of truth"
>>> "{0}".format(s)
# default formatting
'The sword of truth'
Chapter 2. Data Types
, . precision type
< left
> right
^ center
= pad
sign and
digits for
0-pad numbers
+ force
- sign if
“ ”
or - as
prefix ints with 0b, 0o, or 0x
use commas for grouping★
field width
for strings;
of decimal
places for
b, c, d,
n, o, x,
e, E, f,
g, G, n,
Figure 2.6 The general form of a format specification
>>> "{0:25}".format(s)
# minimum width 25
'The sword of truth
>>> "{0:>25}".format(s) # right align, minimum width 25
The sword of truth'
>>> "{0:^25}".format(s) # center align, minimum width 25
The sword of truth
>>> "{0:-^25}".format(s) # - fill, center align, minimum width 25
'---The sword of truth----'
>>> "{0:.<25}".format(s) # . fill, left align, minimum width 25
'The sword of truth.......'
>>> "{0:.10}".format(s) # maximum width 10
'The sword '
In the penultimate example we had to specify the left alignment (even though
this is the default). If we left out the <, we would have :.25, and this simply
means a maximum field width of 25 characters.
As we noted earlier, it is possible to have replacement fields inside format specifications. This makes it possible to have computed formats. Here, for example,
are two ways of setting a string’s maximum width using a maxwidth variable:
>>> maxwidth = 12
>>> "{0}".format(s[:maxwidth])
'The sword of'
>>> "{0:.{1}}".format(s, maxwidth)
'The sword of'
The first approach uses standard string slicing; the second uses an inner
replacement field.
The grouping comma was introduced with Python 3.1.
For integers, the format specification allows us to control the fill character, the
alignment within the field, the sign, whether to use a nonlocale-aware comma
separator to group digits (from Python 3.1), the minimum field width, and the
number base.
An integer format specification begins with a colon, after which we can have
an optional pair of characters—a fill character (which may not be }) and an
alignment character (< for left align, ^ for center, > for right align, and = for the
filling to be done between the sign and the number). Next is an optional sign
character: + forces the output of the sign, - outputs the sign only for negative
numbers, and a space outputs a space for positive numbers and a - sign for
negative numbers. Then comes an optional minimum width integer—this can
be preceded by a # character to get the base prefix output (for binary, octal, and
hexadecimal numbers), and by a 0 to get 0-padding. Then, from Python 3.1,
comes an optional comma—if present this will cause the number’s digits to be
grouped into threes with a comma separating each group. If we want the output in a base other than decimal we must add a type character—b for binary,
o for octal, x for lowercase hexadecimal, and X for uppercase hexadecimal, although for completeness, d for decimal integer is also allowed. There are two
other type characters: c, which means that the Unicode character corresponding to the integer should be output, and n, which outputs numbers in a localesensitive way. (Note that if n is used, using , doesn’t make sense.)
We can get 0-padding in two different ways:
>>> "{0:0=12}".format(8749203)
>>> "{0:0=12}".format(-8749203)
>>> "{0:012}".format(8749203)
>>> "{0:012}".format(-8749203)
# 0 fill, minimum width 12
# 0 fill, minimum width 12
# 0-pad and minimum width 12
# 0-pad and minimum width 12
The first two examples have a fill character of 0 and fill between the sign and
the number itself (=). The second two examples have a minimum width of 12
and 0-padding.
Here are some alignment examples:
>>> "{0:*<15}".format(18340427)
>>> "{0:*>15}".format(18340427)
>>> "{0:*^15}".format(18340427)
>>> "{0:*^15}".format(-18340427)
# * fill, left align, min width 15
# * fill, right align, min width 15
# * fill, center align, min width 15
# * fill, center align, min width 15
Chapter 2. Data Types
Here are some examples that show the effects of the sign characters:
>>> "[{0: }] [{1: }]".format(539802, -539802) # space or - sign
'[ 539802] [-539802]'
>>> "[{0:+}] [{1:+}]".format(539802, -539802) # force sign
'[+539802] [-539802]'
>>> "[{0:-}] [{1:-}]".format(539802, -539802) # - sign if needed
'[539802] [-539802]'
And here are two examples that use some of the type characters:
>>> "{0:b} {0:o} {0:x} {0:X}".format(14613198)
'110111101111101011001110 67575316 deface DEFACE'
>>> "{0:#b} {0:#o} {0:#x} {0:#X}".format(14613198)
'0b110111101111101011001110 0o67575316 0xdeface 0XDEFACE'
It is not possible to specify a maximum field width for integers. This is because
doing so might require digits to be chopped off, thereby rendering the integer
If we are using Python 3.1 and use a comma in the format specification, the
integer will use commas for grouping. For example:
>>> "{0:,}
Both fields have grouping applied, and in addition, the second field is padded
with *s, right aligned, and given a minimum width of 13 characters. This is
very convenient for many scientific and financial programs, but it does not take
into account the current locale. For example, many Continental Europeans
would expect the thousands separator to be . and the decimal separator to
be ,.
The last format character available for integers (and which is also available for
floating-point numbers) is n. This has the same effect as d when given an integer and the same effect as g when given a floating-point number. What makes n
special is that it respects the current locale, and will use the locale-specific decimal separator and grouping separator in the output it produces. The default
locale is called the C locale, and for this the decimal and grouping characters
are a period and an empty string. We can respect the user’s locale by starting
our programs with the following two lines as the first executable statements:★
import locale
locale.setlocale(locale.LC_ALL, "")
In multithreaded programs it is best to call locale.setlocale() only once, at program start-up, and
before any additional threads have been started, since the function is not usually thread-safe.
Passing an empty string as the locale tells Python to try to automatically
determine the user’s locale (e.g., by examining the LANG environment variable),
with a fallback of the C locale. Here are some examples that show the effects
of different locales on an integer and a floating-point number:
x, y = (1234567890, 1234.56)
locale.setlocale(locale.LC_ALL, "C")
c = "{0:n} {1:n}".format(x, y)
# c == "1234567890 1234.56"
locale.setlocale(locale.LC_ALL, "en_US.UTF-8")
en = "{0:n} {1:n}".format(x, y)
# en == "1,234,567,890 1,234.56"
locale.setlocale(locale.LC_ALL, "de_DE.UTF-8")
de = "{0:n} {1:n}".format(x, y)
# de == "1.234.567.890 1.234,56"
Although n is very useful for integers, it is of more limited use with floatingpoint numbers because as soon as they become large they are output using exponential form.
For floating-point numbers, the format specification gives us control over the
fill character, the alignment within the field, the sign, whether to use a nonlocale aware comma separator to group digits (from Python 3.1), the minimum field width, the number of digits after the decimal place, and whether to
present the number in standard or exponential form, or as a percentage.
The format specification for floating-point numbers is the same as for integers,
except for two differences at the end. After the optional minimum width—from
Python 3.1, after the optional grouping comma—we can specify the number of
digits after the decimal place by writing a period followed by an integer. We can
also add a type character at the end: e for exponential form with a lowercase e,
E for exponential form with an uppercase E, f for standard floating-point form,
g for “general” form—this is the same as f unless the number is very large, in
which case it is the same as e—and G, which is almost the same as g, but uses
either f or E. Also available is %—this results in the number being multiplied by
100 with the resultant number output in f format with a % symbol appended.
Here are a few examples that show exponential and standard forms:
>>> amount = (10 ** 3) * math.pi
>>> "[{0:12.2e}] [{0:12.2f}]".format(amount)
3.14e+03] [
>>> "[{0:*>12.2e}] [{0:*>12.2f}]".format(amount)
'[****3.14e+03] [*****3141.59]'
>>> "[{0:*>+12.2e}] [{0:*>+12.2f}]".format(amount)
'[***+3.14e+03] [****+3141.59]'
The first example has a minimum width of 12 characters and has 2 digits after
the decimal point. The second example builds on the first, and adds a * fill
character. If we use a fill character we must also have an alignment character,
so we have specified align right (even though that is the default for numbers).
Chapter 2. Data Types
The third example builds on the previous two, and adds the + sign character to
force the output of the sign.
In Python 3.0, decimal.Decimal numbers are treated by str.format() as strings
rather than as numbers. This makes it quite tricky to get nicely formatted output. From Python 3.1, decimal.Decimal numbers can be formatted as floats, including support for , to get comma-separated groups. Here is an example—we
have omitted the field name since we don’t need it for Python 3.1:
>>> "{:,.6f}".format(decimal.Decimal("1234567890.1234567890"))
If we omitted the f format character (or used the g format character), the
number would be formatted as '1.23457E+9'.
Python 3.0 does not provide any direct support for formatting complex
numbers—support was added with Python 3.1. However, we can easily solve
this by formatting the real and imaginary parts as individual floating-point
numbers. For example:
>>> "{0.real:.3f}{0.imag:+.3f}j".format(4.75917+1.2042j)
>>> "{0.real:.3f}{0.imag:+.3f}j".format(4.75917-1.2042j)
We access each attribute of the complex number individually, and format them
both as floating-point numbers, in this case with three digits after the decimal
place. We have also forced the sign to be output for the imaginary part; we
must add on the j ourselves.
Python 3.1 supports formatting complex numbers using the same syntax as for
>>> "{:,.4f}".format(3.59284e6-8.984327843e6j)
One slight drawback of this approach is that exactly the same formatting is
applied to both the real and the imaginary parts; but we can always use the
Python 3.0 technique of accessing the complex number’s attributes individually if we want to format each one differently.
In the preceding subsubsections we closely examined the str.format() method’s
format specifications, and we have seen many code snippets that show particular aspects. In this subsubsection we will review a small yet useful example
that makes use of str.format() so that we can see format specifications in a
realistic context. The example also uses some of the string methods we saw in
the previous section, and introduces a function from the unicodedata module.★
The program has just 25 lines of executable code. It imports two modules, sys
and unicodedata, and defines one custom function, print_unicode_table(). We’ll
begin by looking at a sample run to see what it does, then we will look at the
code at the end of the program where processing really starts, and finally we
will look at the custom function. spoked
------- ----- --- ---------------------------------------10018
✢ Four Teardrop-Spoked Asterisk
✣ Four Balloon-Spoked Asterisk
✤ Heavy Four Balloon-Spoked Asterisk
✥ Four Club-Spoked Asterisk
✳ Eight Spoked Asterisk
✽ Teardrop-Spoked Asterisk
✼ Open Centre Teardrop-Spoked Asterisk
✽ Heavy Teardrop-Spoked Asterisk
❃ Heavy Teardrop-Spoked Pinwheel Asterisk
❈ Balloon-Spoked Asterisk
❊ Eight Teardrop-Spoked Propeller Asterisk
❋ Heavy Eight Teardrop-Spoked Propeller Asterisk
If run with no arguments, the program produces a table of every Unicode
character, starting from the space character and going up to the character with
the highest available code point. If an argument is given, as in the example,
only those rows in the table where the lowercased Unicode character name
contains the argument are printed.
word = None
if len(sys.argv) > 1:
if sys.argv[1] in ("-h", "--help"):
print("usage: {0} [string]".format(sys.argv[0]))
word = 0
word = sys.argv[1].lower()
if word != 0:
This program assumes that the console uses the Unicode UTF-8 encoding. Unfortunately, the Windows console has poor UTF-8 support. As a workaround, the examples include, a version of the program that writes its output to a file which can then be
opened using a UTF-8-savvy editor, such as IDLE.
Chapter 7
(File Handling)
➤ 287
Chapter 2. Data Types
After the imports and the creation of the print_unicode_table() function, execution reaches the code shown here. We begin by assuming that the user has
not given a word to match on the command line. If a command-line argument
is given and is -h or --help, we print the program’s usage information and set
word to 0 as a flag to indicate that we are finished. Otherwise, we set the word
to a lowercase copy of the argument the user typed in. If the word is not 0, then
we print the table.
When we print the usage information we use a format specification that just
has the format name—in this case, the position number of the argument. We
could have written the line like this instead:
print("usage: {0[0]} [string]".format(sys.argv))
Using this approach the first 0 is the index position of the argument we want
to use, and [0] is the index within the argument, and it works because sys.argv
is a list.
def print_unicode_table(word):
print("------- ----- ---
code = ord(" ")
end = sys.maxunicode
while code < end:
c = chr(code)
name =, "*** unknown ***")
if word is None or word in name.lower():
print("{0:7} {0:5X} {0:^3c} {1}".format(
code, name.title()))
code += 1
We’ve used a couple of blank lines for the sake of clarity. The first two lines of
the function’s suite print the title lines. The first str.format() prints the text
“name” centered in a field 40 characters wide, whereas the second one prints
an empty string in a field 40 characters wide, using a fill character of “-”, and
aligned left. (We must give an alignment if we specify a fill character.) An
alternative approach for the second line is this:
{0}".format("-" * 40))
Here we have used the string replication operator (*) to create a suitable string,
and simply inserted it into the format string. A third alternative would be to
simply type in 40 “-”s and use a literal string.
We keep track of Unicode code points in the code variable, initializing it to
the code point for a space (0x20). We set the end variable to be the highest
Unicode code point available—this will vary depending on whether Python
was compiled to use the UCS-2 or the UCS-4 character encoding.
Inside the while loop we get the Unicode character that corresponds to the code
point using the chr() function. The function returns the
Unicode character name for the given Unicode character; its optional second
argument is the name to use if no character name is defined.
If the user didn’t specify a word (word is None), or if they did and it is in a lowercased copy of the Unicode character name, then we print the corresponding row.
Although we pass the code variable to the str.format() method only once, it is
used three times in the format string, first to print the code as an integer in a
field 7 characters wide (the fill character defaults to space, so we did not need
to specify it), second to print the code as an uppercase hexadecimal number
in a field 5 characters wide, and third to print the Unicode character that
corresponds to the code—using the “c” format specifier, and centered in a field
with a minimum width of three characters. Notice that we did not have to
specify the type “d” in the first format specification; this is because it is the
default for integer arguments. The second argument is the character’s Unicode
character name, printed using “title” case, that is, with the first letter of each
word uppercased, and all other letters lowercased.
Now that we are familiar with the versatile str.format() method, we will make
great use of it throughout the rest of the book.
Character Encodings
Ultimately, computers can store only bytes, that is, 8-bit values which, if unsigned, range from 0x00 to 0xFF. Every character must somehow be represented
in terms of bytes. In the early days of computing the pioneers devised encoding
schemes that assigned a particular character to a particular byte. For example,
using the ASCII encoding, A is represented by 0x41, B by 0x42, and so on. In the
U.S. and Western Europe the Latin-1 encoding was often used; its characters
in the range 0x20–0x7E are the same as the corresponding characters in 7-bit
ASCII, with those in the range 0xA0–0xFF used for accented characters and other symbols needed by those using non-English Latin alphabets. Many other
encodings have been devised over the years, and now there are lots of them in
use—however, development has ceased for many of them, in favor of Unicode.
Having all these different encodings has proved very inconvenient, especially
when writing internationalized software. One solution that has been almost
universally adopted is the Unicode encoding. Unicode assigns every character to an integer—called a code point in Unicode-speak—just like the earlier
encodings. But Unicode is not limited to using one byte per character, and is
therefore able to represent every character in every language in a single
Chapter 2. Data Types
ing, so unlike other encodings, Unicode can handle characters from a mixture
of languages, rather than just one.
But how is Unicode stored? Currently, slightly more than 100 000 Unicode
characters are defined, so even using signed numbers, a 32-bit integer is more
than adequate to store any Unicode code point. So the simplest way to store
Unicode characters is as a sequence of 32-bit integers, one integer per character. This sounds very convenient since it should produce a one to one mapping
of characters to 32-bit integers, which would make indexing to a particular
character very fast. However, in practice things aren’t so simple, since some
Unicode characters can be represented by one or by two code points—for example, é can be represented by the single code point 0xE9 or by two code points,
0x65 and 0x301 (e and a combining acute accent).
Nowadays, Unicode is usually stored both on disk and in memory using UTF8, UTF-16, or UTF-32. The first of these, UTF-8, is backward compatible with
7-bit ASCII since its first 128 code points are represented by single-byte values that are the same as the 7-bit ASCII character values. To represent all the
other Unicode characters, UTF-8 uses two, three, or more bytes per character.
This makes UTF-8 very compact for representing text that is all or mostly English. The Gtk library (used by the GNOME windowing system, among others)
uses UTF-8, and it seems that UTF-8 is becoming the de facto standard format
for storing Unicode text in files—for example, UTF-8 is the default format for
XML, and many web pages these days use UTF-8.
A lot of other software, such as Java, uses UCS-2 (which in modern form is
the same as UTF-16). This representation uses two or four bytes per character,
with the most common characters represented by two bytes. The UTF-32 representation (also called UCS-4) uses four bytes per character. Using UTF-16
or UTF-32 for storing Unicode in files or for sending over a network connection
has a potential pitfall: If the data is sent as integers then the endianness matters. One solution to this is to precede the data with a byte order mark so that
readers can adapt accordingly. This problem doesn’t arise with UTF-8, which
is another reason why it is so popular.
Python represents Unicode using either UCS-2 (UTF-16) format, or UCS-4
(UTF-32) format. In fact, when using UCS-2, Python uses a slightly simplified
version that always uses two bytes per character and so can only represent code
points up to 0xFFFF. When using UCS-4, Python can represent all the Unicode
code points. The maximum code point is stored in the read-only sys.maxunicode
attribute—if its value is 65 535, then Python was compiled to use UCS-2; if
larger, then Python is using UCS-4.
The str.encode() method returns a sequence of bytes—actually a bytes object,
covered in Chapter 7—encoded according to the encoding argument we supply.
Using this method we can get some insight into the difference between encodings, and why making incorrect encoding assumptions can lead to errors:
>>> artist = "Tage Åsén"
>>> artist.encode("Latin1")
b'Tage \xc5s\xe9n'
>>> artist.encode("CP850")
b'Tage \x8fs\x82n'
>>> artist.encode("utf8")
b'Tage \xc3\x85s\xc3\xa9n'
>>> artist.encode("utf16")
b'\xff\xfeT\x00a\x00g\x00e\x00 \x00\xc5\x00s\x00\xe9\x00n\x00'
A b before an opening quote signifies a bytes literal rather than a string
literal. As a convenience, when creating bytes literals we can use a mixture of
printable ASCII characters and hexadecimal escapes.
We cannot encode Tage Åsén’s name using the ASCII encoding because it does
not have the Å character or any accented characters, so attempting to do so
will result in a UnicodeEncodeError exception being raised. The Latin-1 encoding (also known as ISO-8859-1) is an 8-bit encoding that has all the necessary
characters for this name. On the other hand, artist Erno″ Bánk would be less
fortunate since the o″ character is not a Latin-1 character and so could not be
successfully encoded. Both names can be successfully encoded using Unicode encodings, of course. Notice, though, that for UTF-16, the first two bytes
are the byte order mark—these are used by the decoding function to detect
whether the data is big- or little-endian so that it can adapt accordingly.
It is worth noting a couple more points about the str.encode() method. The
first argument (the encoding name) is case-insensitive, and hyphens and underscores in the name are treated as equivalent, so “us-ascii” and “US_ASCII”
are considered the same. There are also many aliases—for example, “latin”,
“latin1”, “latin_1”, “ISO-8859-1”, “CP819”, and some others are all “Latin-1”.
The method can also accept an optional second argument which is used to tell it
how to handle errors. For example, we can encode any string into ASCII if we
pass a second argument of “ignore” or “replace”—at the price of losing data, of
course—or losslessly if we use “backslashreplace” which replaces non-ASCII
characters with \x, \u, and \U escapes. For example, artist.encode("ascii",
"ignore") will produce b'Tage sn' and artist.encode("ascii", "replace") will
produce b'Tage ?s?n', whereas artist.encode("ascii", "backslashreplace")
will produce b'Tage \xc5s\xe9n'. (We can also get an ASCII string using
"{0!a}".format(artist), which produces 'Tage \xc5s\xe9n'.)
The complement of str.encode() is bytes.decode() (and bytearray.decode())
which returns a string with the bytes decoded using the given encoding.
For example:
>>> print(b"Tage \xc3\x85s\xc3\xa9n".decode("utf8"))
Tage Åsén
Chapter 2. Data Types
>>> print(b"Tage \xc5s\xe9n".decode("latin1"))
Tage Åsén
The differences between the 8-bit Latin-1, CP850 (an IBM PC encoding), and
UTF-8 encodings make it clear that guessing encodings is not likely to be a
successful strategy. Fortunately, UTF-8 is becoming the de facto standard for
plain text files, so later generations may not even know that other encodings
ever existed.
Python .py files use UTF-8, so Python always knows the encoding to use with
string literals. This means that we can type any Unicode characters into our
strings—providing our editor supports this.★
When Python reads data from external sources such as sockets, it cannot know
what encoding is used, so it returns bytes which we can then decode accordingly. For text files Python takes a softer approach, using the local encoding unless
we specify an encoding explicitly.
Fortunately, some file formats specify their encoding. For example, we can assume that an XML file uses UTF-8, unless the <?xml?> directive explicitly specifies a different encoding. So when reading XML we might extract, say, the first
1 000 bytes, look for an encoding specification, and if found, decode the file using the specified encoding, otherwise falling back to decoding using UTF-8. This
approach should work for any XML or plain text file that uses any of the single byte encodings supported by Python, except for EBCDIC-based encodings
(CP424, CP500) and a few others (CP037, CP864, CP865, CP1026, CP1140, HZ,
SHIFT-JIS-2004, SHIFT-JISX0213). Unfortunately, this approach won’t work
for multibyte encodings (such as UTF-16 and UTF-32). At least two Python
packages for automatically detecting a file’s encoding are available from the
Python Package Index,
In this section we will draw on what we have covered in this chapter and the
one before, to present two small but complete programs to help consolidate
what we have learned so far. The first program is a bit mathematical, but it is
quite short at around 35 lines. The second is concerned with text processing
and is more substantial, with seven functions in around 80 lines of code.
Quadratic equations are equations of the form ax2 + bx + c = 0 where a ≠ 0
describe parabolas. The roots of such equations are derived from the formula
It is possible to use other encodings. See the Python Tutorial’s “Source Code Encoding” topic.
x = −b± √2a −4ac. The b − 4ac part of the formula is called the discriminant—if it
is positive there are two real roots, if it is zero there is one real root, and if it is
negative there are two complex roots. We will write a program that accepts the
a, b, and c factors from the user (with the b and c factors allowed to be 0), and
then calculates and outputs the root or roots.★
First we will look at a sample run, and then we will review the code.
ax² + bx + c = 0
enter a: 2.5
enter b: 0
enter c: -7.25
2.5x² + 0.0x + -7.25 = 0 → x = 1.70293863659 or x = -1.70293863659
With factors 1.5, -3, and 6, the output (with some digits trimmed) is:
1.5x² + -3.0x + 6.0 = 0 → x = (1+1.7320508j) or x = (1-1.7320508j)
The output isn’t quite as tidy as we’d like—for example, rather than + -3.0x
it would be nicer to have - 3.0x, and we would prefer not to have any 0 factors
shown at all. You will get the chance to fix these problems in the exercises.
Now we will turn to the code, which begins with three imports:
import cmath
import math
import sys
We need both the float and the complex math libraries since the square root
functions for real and complex numbers are different, and we need sys for
sys.float_info.epsilon which we need to compare floating-point numbers
with 0.
We also need a function that can get a floating-point number from the user:
def get_float(msg, allow_zero):
x = None
while x is None:
x = float(input(msg))
if not allow_zero and abs(x) < sys.float_info.epsilon:
print("zero is not allowed")
x = None
Since the Windows console has poor UTF-8 support, there are problems with a couple of the
characters (² and →) that uses. We have provided which displays the
correct symbols on Linux and Mac OS X, and alternatives (^2 and ->) on Windows.
Chapter 2. Data Types
except ValueError as err:
return x
This function will loop until the user enters a valid floating-point number (such
as 0.5, -9, 21, 4.92), and will accept 0 only if allow_zero is True.
Once the get_float() function is defined, the rest of the code is executed. We’ll
look at it in three parts, starting with the user interaction:
a = get_float("enter a:
b = get_float("enter b:
c = get_float("enter c:
TWO} + bx + c = 0")
", False)
", True)
", True)
Thanks to the get_float() function, getting the a, b, and c factors is simple. The
Boolean second argument says whether 0 is acceptable.
x1 = None
x2 = None
discriminant = (b ** 2) - (4 * a * c)
if discriminant == 0:
x1 = -(b / (2 * a))
if discriminant > 0:
root = math.sqrt(discriminant)
else: # discriminant < 0
root = cmath.sqrt(discriminant)
x1 = (-b + root) / (2 * a)
x2 = (-b - root) / (2 * a)
The code looks a bit different to the formula because we begin by calculating
the discriminant. If the discriminant is 0, we know that we have one real
solution and so we calculate it directly. Otherwise, we take the real or complex
square root of the discriminant and calculate the two roots.
equation = ("{0}x\N{SUPERSCRIPT TWO} + {1}x + {2} = 0"
" \N{RIGHTWARDS ARROW} x = {3}").format(a, b, c, x1)
if x2 is not None:
equation += " or x = {0}".format(x2)
We haven’t done any fancy formatting since Python’s defaults for floating-point
numbers are fine for this example, but we have used Unicode character names
for a couple of special characters.
Using str.
mapping unpacking
81 ➤
A more robust alternative to using positional arguments with their index positions as field names, is to use the dictionary returned by locals(), a technique
we saw earlier in the chapter.
equation = ("{a}x\N{SUPERSCRIPT TWO} + {b}x + {c} = 0"
" \N{RIGHTWARDS ARROW} x = {x1}").format(**locals())
And if we are using Python 3.1, we could omit the field names and leave Python
to populate the fields using the positional arguments passed to str.format().
equation = ("{}x\N{SUPERSCRIPT TWO} + {}x + {} = 0"
" \N{RIGHTWARDS ARROW} x = {}").format(a, b, c, x1)
This is convenient, but not as robust as using named parameters, nor as
versatile if we needed to use format specifications. Nonetheless, for many
simple cases this syntax is both easy and useful.
One common requirement is to take a data set and present it using HTML. In
this subsection we will develop a program that reads a file that uses a simple
CSV (Comma Separated Value) format and outputs an HTML table containing
the file’s data. Python comes with a powerful and sophisticated module for
handling CSV and similar formats—the csv module—but here we will write
all the code by hand.
The CSV format we will support has one record per line, with each record
divided into fields by commas. Each field can be either a string or a number.
Strings must be enclosed in single or double quotes and numbers should be
unquoted unless they contain commas. Commas are allowed inside strings,
and must not be treated as field separators. We assume that the first record
contains field labels. The output we will produce is an HTML table with text
left-aligned (the default in HTML) and numbers right-aligned, with one row
per record and one cell per field.
The program must output the HTML table’s opening tag, then read each line of
data and for each one output an HTML row, and at the end output the HTML
table’s closing tag. We want the background color of the first row (which will
display the field labels) to be light green, and the background of the data rows
to alternate between white and light yellow. We must also make sure that the
special HTML characters (“&”, “<”, and “>”) are properly escaped, and we want
strings to be tidied up a bit.
Here’s a tiny piece of sample data:
Chapter 2. Data Types
"BAHAMAS, THE",1,1,1,1,1
Assuming the sample data is in the file data/co2-sample.csv, and given
the command < data/co2-sample.csv > co2-sample.html, the file
co2-sample.html will have contents similar to this:
<table border='1'><tr bgcolor='lightgreen'>
<td>Country</td><td align='right'>2000</td><td align='right'>2001</td>
<td align='right'>2002</td><td align='right'>2003</td>
<td align='right'>2004</td></tr>
<tr bgcolor='lightyellow'><td>Argentina</td>
<td align='right'>37</td><td align='right'>35</td>
<td align='right'>33</td><td align='right'>36</td>
<td align='right'>39</td></tr>
We’ve tidied the output slightly and omitted some lines where indicated by
ellipses. We have used a very simple version of HTML—HTML 4 transitional,
with no style sheet. Figure 2.7 shows what the output looks like in a web
Figure 2.7 A table in a web browser
Now that we’ve seen how the program is used and what it does, we are ready
to review the code. The program begins with the import of the sys module; we
won’t show this, or any other imports from now on, unless they are unusual
or warrant discussion. And the last statement in the program is a single
function call:
Although Python does not need an entry point as some languages require, it
is quite common in Python programs to create a function called main() and to
call it to start off processing. Since no function can be called before it has been
created, we must make sure we call main() after the functions it relies on have
been defined. The order in which the functions appear in the file (i.e., the order
in which they are created) does not matter.
In the program, the first function we call is main() which in turn
calls print_start() and then print_line(). And print_line() calls extract_
fields() and escape_html(). The program structure we have used is shown in
Figure 2.8.
import sys
def main():
def print_start():
def print_line():
def extract_fields():
def escape_html():
def print_end():
Figure 2.8 The program’s structure
When Python reads a file it begins at the top. So for this example, it starts by
performing the import, then it creates the main() function, and then it creates
the other functions in the order in which they appear in the file. When Python
finally reaches the call to main() at the end of the file, all the functions that
main() will call (and all the functions that those functions will call) now exist.
Execution as we normally think of it begins where the call to main() is made.
We will look at each function in turn, starting with main().
def main():
maxwidth = 100
count = 0
while True:
line = input()
if count == 0:
color = "lightgreen"
elif count % 2:
color = "white"
color = "lightyellow"
Chapter 2. Data Types
print_line(line, color, maxwidth)
count += 1
except EOFError:
The maxwidth variable is used to constrain the number of characters in a
cell—if a field is bigger than this we will truncate it and signify this by adding
an ellipsis to the truncated text. We’ll look at the print_start(), print_line(),
and print_end() functions in a moment. The while loop iterates over each line
of input—this could come from the user typing at the keyboard, but we expect
it to be a redirected file. We set the color we want to use and call print_line()
to output the line as an HTML table row.
def print_start():
print("<table border='1'>")
def print_end():
We could have avoided creating these two functions and simply put the relevant print() function calls in main(). But we prefer to separate out the logic
since this is more flexible, even though it doesn’t really matter in this small
def print_line(line, color, maxwidth):
print("<tr bgcolor='{0}'>".format(color))
fields = extract_fields(line)
for field in fields:
if not field:
number = field.replace(",", "")
x = float(number)
print("<td align='right'>{0:d}</td>".format(round(x)))
except ValueError:
field = field.title()
field = field.replace(" And ", " and ")
if len(field) <= maxwidth:
field = escape_html(field)
field = "{0} ...".format(
We cannot use str.split(",") to split each line into fields because commas
can occur inside quoted strings. So we have farmed this work out to the
extract_fields() function. Once we have a list of the fields (as strings, with no
surrounding quotes), we iterate over them, creating a table cell for each one.
If a field is empty, we output an empty cell. If a field is quoted, it could be
a string or it could be a number that has been quoted to allow for internal
commas, for example, "1,566". To account for this, we make a copy of the field
with commas removed and try to convert the field to a float. If the conversion is
successful we output a right-aligned cell with the field rounded to the nearest
whole number and output it as an integer. If the conversion fails we output the
field as a string. In this case we use str.title() to neaten the case of the letters
and we replace the word And with and as a correction to str.title()’s effect.
If the field isn’t too long we use all of it, otherwise we truncate it to maxwidth
characters and add an ellipsis to signify the truncation, and in either case we
escape any special HTML characters the field might contain.
def extract_fields(line):
fields = []
field = ""
quote = None
for c in line:
if c in "\"'":
if quote is None: # start of quoted string
quote = c
elif quote == c: # end of quoted string
quote = None
field += c
# other quote inside quoted string
if quote is None and c == ",": # end of a field
field = ""
field += c
# accumulating a field
if field:
fields.append(field) # adding the last field
return fields
This function reads the line it is given character by character, accumulating
a list of fields—each one a string without any enclosing quotes. The function
copes with fields that are unquoted, and with fields that are quoted with single
or double quotes, and correctly handles commas and quotes (single quotes in
double quoted strings, double quotes in single quoted strings).
Chapter 2. Data Types
def escape_html(text):
text = text.replace("&", "&amp;")
text = text.replace("<", "&lt;")
text = text.replace(">", "&gt;")
return text
This function straightforwardly replaces each special HTML character with
the appropriate HTML entity. We must of course replace ampersands first,
although the order doesn’t matter for the angle brackets. Python’s standard
library includes a slightly more sophisticated version of this function—you’ll
get the chance to use it in the exercises, and will see it again in Chapter 7.
This chapter began by showing the list of Python’s keywords and described the
rules that Python applies to identifiers. Thanks to Python’s Unicode support,
identifiers are not limited to a subset of characters from a small character set
like ASCII or Latin-1.
We also described Python’s int data type, which differs from similar types in
most other languages in that it has no intrinsic size limitation. Python integers
can be as large as the machine’s memory will allow, and it is perfectly feasible to
work with numbers that are hundreds of digits long. All of Python’s most basic
data types are immutable, but this is rarely noticable since the augmented assignment operators (+=, *=, -=, /=, and others) means that we can use a very natural syntax while behind the scenes Python creates result objects and rebinds
our variables to them. Literal integers are usually written as decimal numbers,
but we can write binary literals using the 0b prefix, octal literals using the 0o
prefix, and hexadecimal literals using the 0x prefix.
When two integers are divided using /, the result is always a float. This is
different from many other widely used languages, but helps to avoid some
quite subtle bugs that can occur when division silently truncates. (And if we
want integer division we can use the // operator.)
Python has a bool data type which can hold either True or False. Python has
three logical operators, and, or, and not, of which the two binary operators (and
and or) use short-circuit logic.
Three kinds of floating-point numbers are available: float, complex, and decimal.Decimal. The most commonly used is float; this is a double-precision
floating-point number whose exact numerical characteristics depend on the
underlying C, C#, or Java library that Python was built with. Complex numbers are represented as two floats, one holding the real value and the other the
imaginary value. The decimal.Decimal type is provided by the decimal module.
These numbers default to having 28 decimal places of accuracy, but this can be
increased or decreased to suit our needs.
All three floating-point types can be used with the appropriate built-in mathematical operators and functions. And in addition, the math module provides a
variety of trigonometric, hyperbolic, and logarithmic functions that can be used
with floats, and the cmath module provides a similar set of functions for complex
Most of the chapter was devoted to strings. Python string literals can be
created using single quotes or double quotes, or using a triple quoted string
if we want to include newlines and quotes without formality. Various escape
sequences can be used to insert special characters such as tab (\t) and newline
(\n), and Unicode characters both using hexadecimal escapes and Unicode
character names. Although strings support the same comparison operators
as other Python types, we noted that sorting strings that contain non-English
characters can be problematic.
Since strings are sequences, the slicing operator ([]) can be used to slice and
stride strings with a very simple yet powerful syntax. Strings can also be
concatenated with the + operator and replicated with the * operator, and we
can also use the augmented assignment versions of these operators (+= and
*=), although the str.join() method is more commonly used for concatenation.
Strings have many other methods, including some for testing string properties
(e.g., str.isspace() and str.isalpha()), some for changing case (e.g., str.lower()
and str.title()), some for searching (e.g., str.find() and str.index()), and
many others.
Python’s string support is really excellent, enabling us to easily find and
extract or compare whole strings or parts of strings, to replace characters or
substrings, and to split strings into a list of substrings and to join lists of
strings into a single string.
Probably the most versatile string method is str.format(). This method is used
to create strings using replacement fields and variables to go in those fields, and
format specifications to precisely define the characteristics of each field which
is replaced with a value. The replacement field name syntax allows us to access
the method’s arguments by position or by name (for keyword arguments), and
to use an index, key, or attribute name to access an argument item or attribute.
The format specifications allow us to specify the fill character, the alignment,
and the minimum field width. Furthermore, for numbers we can also control
how the sign is output, and for floating-point numbers we can specify the number of digits after the decimal point and whether to use standard or exponential notation.
We also discussed the thorny issue of character encodings. Python .py files use
the Unicode UTF-8 encoding by default and so can have comments, identifiers,
and data written in just about any human language. We can convert a string
Chapter 2. Data Types
into a sequence of bytes using a particular encoding using the str.encode()
method, and we can convert a sequence of bytes that use a particular encoding
back to a string using the bytes.decode() method. The wide variety of character encodings currently in use can be very inconvenient, but UTF-8 is fast becoming the de facto standard for plain text files (and is already the default for
XML files), so this problem should diminish in the coming years.
In addition to the data types covered in this chapter, Python provides two other
built-in data types, bytes and bytearray, both of which are covered in Chapter 7.
Python also provides several collection data types, some built-in and others
in the standard library. In the next chapter we will look at Python’s most
important collection data types.
1. Modify the program so that the user can enter several
separate words on the command line, and print rows only where the
Unicode character name contains all the words the user has specified.
This means that we can type commands like this: greek symbol
One way of doing this is to replace the word variable (which held 0, None,
or a string), with a words list. Don’t forget to update the usage information as well as the code. The changes involve adding less than ten lines
of code, and changing less than ten more. A solution is provided in file (Windows and cross-platform users should modify; a solution is provided in
2. Modify so that 0.0 factors are not output, and so that negative
factors are output as - n rather than as + -n. This involves replacing the
last five lines with about fifteen lines. A solution is provided in (Windows and cross-platform users should modify; a solution is provided in
3. Delete the escape_html() function from, and use the xml.sax.
saxutils.escape() function from the xml.sax.saxutils module instead. This
is easy, requiring one new line (the import), five deleted lines (the unwanted function), and one changed line (to use xml.sax.saxutils.escape() instead of escape_html()). A solution is provided in
4. Modify again, this time adding a new function called process_options(). This function should be called from main() and should
return a tuple of two values: maxwidth (an int) and format (a str). When
process_options() is called it should set a default maxwidth of 100, and a
default format of “.0f”—this will be used as the format specifier when outputting numbers.
If the user has typed “-h” or “--help” on the command line, a usage message
should be output and (None, None) returned. (In this case main() should
do nothing.) Otherwise, the function should read any command-line
arguments that are given and perform the appropriate assignments. For
example, setting maxwidth if “maxwidth=n” is given, and similarly setting
format if “format=s” is given. Here is a run showing the usage output: -h
usage: [maxwidth=int] [format=str] < infile.csv > outfile.html
maxwidth is an optional integer; if specified, it sets the maximum
number of characters that can be output for string fields,
otherwise a default of 100 characters is used.
format is the format to use for numbers; if not specified it
defaults to ".0f".
And here is a command line with both options set: maxwidth=20 format=0.2f < mydata.csv > mydata.html
Don’t forget to modify print_line() to make use of the format for outputting numbers—you’ll need to pass in an extra argument, add one line,
and modify another line. And this will slightly affect main() too. The process_options() function should be about twenty-five lines (including about
nine for the usage message). This exercise may prove challenging for inexperienced programmers.
Two files of test data are provided: data/co2-sample.csv and data/co2-fromfossilfuels.csv. A solution is provided in In Chapter 5
we will see how to use Python’s optparse module to simplify command-line
● Sequence Types
● Set Types
● Mapping Types
● Iterating and Copying Collections
Collection Data Types
In the preceding chapter we learned about Python’s most important fundamental data types. In this chapter we will extend our programming options
by learning how to gather data items together using Python’s collection data
types. We will cover tuples and lists, and also introduce new collection data
types, including sets and dictionaries, and cover all of them in depth.★
In addition to collections, we will also see how to create data items that are
aggregates of other data items (like C or C++ structs or Pascal records)—such
items can be treated as a single unit when this is convenient for us, while
the items they contain remain individually accessible. Naturally, we can put
aggregated items in collections just like any other items.
Having data items in collections makes it much easier to perform operations
that must be applied to all of the items, and also makes it easier to handle collections of items read in from files. We’ll cover the very basics of text file handling in this chapter as we need them, deferring most of the detail (including
error handling) to Chapter 7.
After covering the individual collection data types, we will look at how to iterate over collections, since the same syntax is used for all of Python’s collections, and we will also explore the issues and techniques involved in copying
Sequence Types
A sequence type is one that supports the membership operator (in), the size
function (len()), slices ([]), and is iterable. Python provides five built-in sequence types: bytearray, bytes, list, str, and tuple—the first two are covered
The definitions of what constitutes a sequence type, a set type, or a mapping type given in this
chapter are practical but informal. More formal definitions are given in Chapter 8.
Chapter 3. Collection Data Types
separately in Chapter 7. Some other sequence types are provided in the standard library, most notably, collections.namedtuple. When iterated, all of these
sequences provide their items in order.
65 ➤
We covered strings in the preceding chapter. In this section we will cover
tuples, named tuples, and lists.
69 ➤
A tuple is an ordered sequence of zero or more object references. Tuples
support the same slicing and striding syntax as strings. This makes it easy to
extract items from a tuple. Like strings, tuples are immutable, so we cannot
replace or delete any of their items. If we want to be able to modify an ordered
sequence, we simply use a list instead of a tuple; or if we already have a tuple
but want to modify it, we can convert it to a list using the list() conversion
function and then apply the changes to the resultant list.
The tuple data type can be called as a function, tuple()—with no arguments
it returns an empty tuple, with a tuple argument it returns a shallow copy of
the argument, and with any other argument it attempts to convert the given
object to a tuple. It does not accept more than one argument. Tuples can also
be created without using the tuple() function. An empty tuple is created using
empty parentheses, (), and a tuple of one or more items can be created by using
commas. Sometimes tuples must be enclosed in parentheses to avoid syntactic
ambiguity. For example, to pass the tuple 1, 2, 3 to a function, we would write
function((1, 2, 3)).
Figure 3.1 shows the tuple t = "venus", -28, "green", "21", 19.74, and the index
positions of the items inside the tuple. Strings are indexed in the same way,
but whereas strings have a character at every position, tuples have an object
reference at each position.
Figure 3.1 Tuple index positions
Tuples provide just two methods, t.count(x), which returns the number of
times object x occurs in tuple t, and t.index(x), which returns the index position
of the leftmost occurrence of object x in tuple t—or raises a ValueError exception if there is no x in the tuple. (These methods are also available for lists.)
In addition, tuples can be used with the operators + (concatenation), * (replication), and [] (slice), and with in and not in to test for membership. The +=
and *= augmented assignment operators can be used even though tuples are
➤ 146
Sequence Types
immutable—behind the scenes Python creates a new tuple to hold the result
and sets the left-hand object reference to refer to it; the same technique is used
when these operators are applied to strings. Tuples can be compared using the
standard comparison operators (<, <=, ==, !=, >=, >), with the comparisons being
applied item by item (and recursively for nested items such as tuples inside
Let’s look at a few slicing examples, starting with extracting one item, and a
slice of items:
>>> hair = "black", "brown", "blonde", "red"
>>> hair[2]
>>> hair[-3:] # same as: hair[1:]
('brown', 'blonde', 'red')
These work the same for strings, lists, and any other sequence type.
>>> hair[:2], "gray", hair[2:]
(('black', 'brown'), 'gray', ('blonde', 'red'))
Here we tried to create a new 5-tuple, but ended up with a 3-tuple that contains
two 2-tuples. This happened because we used the comma operator with three
items (a tuple, a string, and a tuple). To get a single tuple with all the items we
must concatenate tuples:
>>> hair[:2] + ("gray",) + hair[2:]
('black', 'brown', 'gray', 'blonde', 'red')
To make a 1-tuple the comma is essential, but in this case, if we had just put
in the comma we would get a TypeError (since Python would think we were
trying to concatenate a string and a tuple), so here we must have the comma
and parentheses.
In this book (from this point on), we will use a particular coding style when
writing tuples. When we have tuples on the left-hand side of a binary operator
or on the right-hand side of a unary statement, we will omit the parentheses,
and in all other cases we will use parentheses. Here are a few examples:
a, b = (1, 2)
# left of binary operator
del a, b
# right of unary statement
def f(x):
return x, x ** 2
# right of unary statement
for x, y in ((1, 1), (2, 4), (3, 9)):
print(x, y)
# left of binary operator
Chapter 3. Collection Data Types
There is no obligation to follow this coding style; some programmers prefer to
always use parentheses—which is the same as the tuple representational form,
whereas others use them only if they are strictly necessary.
>>> eyes = ("brown", "hazel", "amber", "green", "blue", "gray")
>>> colors = (hair, eyes)
>>> colors[1][3:-1]
('green', 'blue')
Here we have nested two tuples inside another tuple. Nested collections to any
level of depth can be created like this without formality. The slice operator []
can be applied to a slice, with as many used as necessary. For example:
>>> things = (1, -7.5, ("pea", (5, "Xyz"), "queue"))
>>> things[2][1][1][2]
Let’s look at this piece by piece, beginning with things[2] which gives us the
third item in the tuple (since the first item has index 0), which is itself a tuple, ("pea", (5, "Xyz"), "queue"). The expression things[2][1] gives us the
second item in the things[2] tuple, which is again a tuple, (5, "Xyz"). And
things[2][1][1] gives us the second item in this tuple, which is the string "Xyz".
Finally, things[2][1][1][2] gives us the third item (character) in the string, that
is, "z".
Tuples are able to hold any items of any data type, including collection types
such as tuples and lists, since what they really hold are object references.
Using complex nested data structures like this can easily become confusing.
One solution is to give names to particular index positions. For example:
aircraft = ("Airbus", "A320-200", (100, 220))
This is certainly more meaningful than writing aircraft[2][1], but it involves
creating lots of variables and is rather ugly. We will see an alternative in the
next subsection.
In the first two lines of the “aircraft” code snippet, we assigned to tuples in
both statements. When we have a sequence on the right-hand side of an
assignment (here we have tuples), and we have a tuple on the left-hand side,
we say that the right-hand side has been unpacked. Sequence unpacking can
be used to swap values, for example:
a, b = (b, a)
Sequence Types
Strictly speaking, the parentheses are not needed on the right, but as we noted
earlier, the coding style used in this book is to omit parentheses for left-hand
operands of binary operators and right-hand operands of unary statements,
but to use parentheses in all other cases.
We have already seen examples of sequence unpacking in the context of for …
in loops. Here is a reminder:
for x, y in ((-3, 4), (5, 12), (28, -45)):
print(math.hypot(x, y))
Here we loop over a tuple of 2-tuples, unpacking each 2-tuple into variables x
and y.
Named Tuples
A named tuple behaves just like a plain tuple, and has the same performance
characteristics. What it adds is the ability to refer to items in the tuple by
name as well as by index position, and this allows us to create aggregates of
data items.
The collections module provides the namedtuple() function. This function is
used to create custom tuple data types. For example:
Sale = collections.namedtuple("Sale",
"productid customerid date quantity price")
The first argument to collections.namedtuple() is the name of the custom tuple
data type that we want to be created. The second argument is a string of spaceseparated names, one for each item that our custom tuples will take. The first
argument, and the names in the second argument, must all be valid Python
identifiers. The function returns a custom class (data type) that can be used
to create named tuples. So, in this case, we can treat Sale just like any other
Python class (such as tuple), and create objects of type Sale. (In object-oriented
terms, every class created this way is a subclass of tuple; object-oriented programming, including subclassing, is covered in Chapter 6.)
Here is an example:
sales = []
sales.append(Sale(432, 921, "2008-09-14", 3, 7.99))
sales.append(Sale(419, 874, "2008-09-15", 1, 18.49))
Here we have created a list of two Sale items, that is, of two custom tuples. We
can refer to items in the tuples using index positions—for example, the price of
the first sale item is sales[0][-1] (i.e., 7.99)—but we can also use names, which
makes things much clearer:
Chapter 3. Collection Data Types
total = 0
for sale in sales:
total += sale.quantity * sale.price
print("Total ${0:.2f}".format(total)) # prints: Total $42.46
The clarity and convenience that named tuples provide are often useful. For
example, here is the “aircraft” example from the previous subsection (110 ➤)
done the nice way:
Aircraft = collections.namedtuple("Aircraft",
"manufacturer model seating")
Seating = collections.namedtuple("Seating", "minimum maximum")
aircraft = Aircraft("Airbus", "A320-200", Seating(100, 220))
When it comes to extracting named tuple items for use in strings there are
three main approaches we can take.
>>> print("{0} {1}".format(aircraft.manufacturer, aircraft.model))
Airbus A320-200
Here we have accessed each of the tuple’s items that we are interested in
using named tuple attribute access. This gives us the shortest and simplest
format string. (And in Python 3.1 we could reduce this format string to just
"{} {}".) But this approach means that we must look at the arguments passed
to str.format() to see what the replacement texts will be. This seems less clear
than using named fields in the format string.
"{0.manufacturer} {0.model}".format(aircraft)
Here we have used a single positional argument and used named tuple attribute names as field names in the format string. This is much clearer than
just using positional arguments alone, but it is a pity that we must specify the positional value (even when using Python 3.1). Fortunately, there is a
nicer way.
Named tuples have a few private methods—that is, methods whose name
begins with a leading underscore. One of them—namedtuple._asdict()—is so
useful that we will show it in action.★
"{manufacturer} {model}".format(**aircraft._asdict())
Using str.
mapping unpacking
81 ➤
The private namedtuple._asdict() method returns a mapping of key–value
pairs, where each key is the name of a tuple element and each value is the cor★
Private methods such as namedtuple._asdict() are not guaranteed to be available in all Python 3.x
versions; although the namedtuple._asdict() method is available in both Python 3.0 and 3.1.
Sequence Types
responding value. We have used mapping unpacking to convert the mapping
into key–value arguments for the str.format() method.
Although named tuples can be very convenient, in Chapter 6 we introduce
object-oriented programming, and there we will go beyond simple named
tuples and learn how to create custom data types that hold data items and that
also have their own custom methods.
69 ➤
A list is an ordered sequence of zero or more object references. Lists support
the same slicing and striding syntax as strings and tuples. This makes it easy
to extract items from a list. Unlike strings and tuples, lists are mutable, so we
can replace and delete any of their items. It is also possible to insert, replace,
and delete slices of lists.
The list data type can be called as a function, list()—with no arguments it
returns an empty list, with a list argument it returns a shallow copy of the
argument, and with any other argument it attempts to convert the given object
to a list. It does not accept more than one argument. Lists can also be created
without using the list() function. An empty list is created using empty brackets, [], and a list of one or more items can be created by using a comma-separated sequence of items inside brackets. Another way of creating lists is to use
a list comprehension—a topic we will cover later in this subsection.
Since all the items in a list are really object references, lists, like tuples, can
hold items of any data type, including collection types such as lists and tuples.
Lists can be compared using the standard comparison operators (<, <=, ==, !=, >=,
>), with the comparisons being applied item by item (and recursively for nested
items such as lists or tuples inside lists).
➤ 118
Given the assignment L = [-17.5, "kilo", 49, "V", ["ram", 5, "echo"], 7], we
get the list shown in Figure 3.2.
['ram', 5, 'echo']
Figure 3.2 List index positions
And given this list, L, we can use the slice operator—repeatedly if necessary—to access items in the list, as the following equalities show:
L[0] == L[-6] == -17.5
L[1] == L[-5] == 'kilo'
L[1][0] == L[-5][0] == 'k'
➤ 146
Chapter 3. Collection Data Types
L[4][2] == L[4][-1] == L[-2][2] == L[-2][-1] == 'echo'
L[4][2][1] == L[4][2][-3] == L[-2][-1][1] == L[-2][-1][-3] == 'c'
Lists can be nested, iterated over, and sliced, the same as tuples. In fact, all
the tuple examples presented in the preceding subsection would work exactly
the same if we used lists instead of tuples. Lists support membership testing
with in and not in, concatenation with +, extending with += (i.e., the appending
of all the items in the right-hand operand), and replication with * and *=. Lists
can also be used with the built-in len() function, and with the del statement
discussed here and described in the sidebar “Deleting Items Using the del
Statement” (➤ 116). In addition, lists provide the methods shown in Table 3.1.
Although we can use the slice operator to access items in a list, in some situations we want to take two or more pieces of a list in one go. This can be done
by sequence unpacking. Any iterable (lists, tuples, etc.) can be unpacked using
the sequence unpacking operator, an asterisk or star (*). When used with two or
more variables on the left-hand side of an assignment, one of which is preceded
by *, items are assigned to the variables, with all those left over assigned to the
starred variable. Here are some examples:
>>> first, *rest = [9, 2, -4, 8, 7]
>>> first, rest
(9, [2, -4, 8, 7])
>>> first, *mid, last = "Charles Philip Arthur George Windsor".split()
>>> first, mid, last
('Charles', ['Philip', 'Arthur', 'George'], 'Windsor')
>>> *directories, executable = "/usr/local/bin/gvim".split("/")
>>> directories, executable
(['', 'usr', 'local', 'bin'], 'gvim')
When the sequence unpacking operator is used like this, the expression *rest,
and similar expressions, are called starred expressions.
Python also has a related concept called starred arguments. For example, if we
have the following function that requires three arguments:
def product(a, b, c):
return a * b * c # here, * is the multiplication operator
we can call it with three arguments, or by using starred arguments:
product(2, 3, 5)
L = [2, 3, 5]
product(2, *L[1:])
Sequence Types
Table 3.1 List Methods
Appends item x to the end of list L
Returns the number of times item x occurs in list L
L += m
Appends all of iterable m’s items to the end of list L; the
operator += does the same thing
Returns the index position of the leftmost occurrence of
item x in list L (or in the start:end slice of L); otherwise,
raises a ValueError exception
L.insert(i, x)
Inserts item x into list L at index position int i
Returns and removes the rightmost item of list L
Returns and removes the item at index position int i in L
Removes the leftmost occurrence of item x from list L, or
raises a ValueError exception if x is not found
Reverses list L in-place
Sorts list L in-place; this method accepts the same key and
reverse optional arguments as the built-in sorted()
➤ 140,
In the first call we provide the three arguments normally. In the second call
we use a starred argument—what happens here is that the three-item list is
unpacked by the * operator, so as far as the function is concerned it has received
the three arguments it is expecting. We could have achieved the same thing
using a 3-tuple. And in the third call we pass the first argument conventionally,
and the other two arguments by unpacking a two-item slice of the L list. Functions and argument passing are covered fully in Chapter 4.
There is never any syntactic ambiguity regarding whether operator * is the
multiplication or the sequence unpacking operator. When it appears on the
left-hand side of an assignment it is the unpacking operator, and when it
appears elsewhere (e.g., in a function call) it is the unpacking operator when
used as a unary operator and the multiplication operator when used as a
binary operator.
We have already seen that we can iterate over the items in a list using the
syntax for item in L:. If we want to change the items in a list the idiom to
use is:
for i in range(len(L)):
L[i] = process(L[i])
The built-in range() function returns an iterator that provides integers. With
one integer argument, n, the iterator range() returns, producing 0, 1, …, n - 1.
➤ 141
Chapter 3. Collection Data Types
Deleting Items Using the del Statement
Although the name of the del statement is reminiscent of the word delete,
it does not necessarily delete any data. When applied to an object reference
that refers to a data item that is not a collection, the del statement unbinds
the object reference from the data item and deletes the object reference.
For example:
>>> x = 8143 # object ref. 'x' created; int of value 8143 created
>>> x
>>> del x # object ref. 'x' deleted; int ready for garbage collection
>>> x
Traceback (most recent call last):
NameError: name 'x' is not defined
When an object reference is deleted, Python schedules the data item to
which it referred to be garbage-collected if no other object references refer to
the data item. When, or even if, garbage collection takes place may be nondeterministic (depending on the Python implementation), so if any cleanup is
required we must handle it ourselves. Python provides two solutions to the
nondeterminism. One is to use a try … finally block to ensure that cleanup
is done, and another is to use a with statement as we will see in Chapter 8.
When del is used on a collection data type such as a tuple or a list, only the
object reference to the collection is deleted. The collection and its items (and
for those items that are themselves collections, for their items, recursively)
are scheduled for garbage collection if no other object references refer to
the collection.
For mutable collections such as lists, del can be applied to individual items
or slices—in both cases using the slice operator, []. If the item or items
referred to are removed from the collection, and if there are no other object
references referring to them, they are scheduled for garbage collection.
We could use this technique to increment all the numbers in a list of integers.
For example:
for i in range(len(numbers)):
numbers[i] += 1
Since lists support slicing, in several cases the same effect can be achieved
using either slicing or one of the list methods. For example, given the list woods
= ["Cedar", "Yew", "Fir"], we can extend the list in either of two ways:
woods += ["Kauri", "Larch"]
woods.extend(["Kauri", "Larch"])
Sequence Types
In either case the result is the list ['Cedar', 'Yew', 'Fir', 'Kauri', 'Larch'].
Individual items can be added at the end of a list using list.append(). Items
can be inserted at any index position within the list using list.insert(), or by
assigning to a slice of length 0. For example, given the list woods = ["Cedar",
"Yew", "Fir", "Spruce"], we can insert a new item at index position 2 (i.e., as
the list’s third item) in either of two ways:
woods[2:2] = ["Pine"]
woods.insert(2, "Pine")
In both cases the result is the list ['Cedar', 'Yew', 'Pine', 'Fir', 'Spruce'].
Individual items can be replaced in a list by assigning to a particular index
position, for example, woods[2] = "Redwood". Entire slices can be replaced by
assigning an iterable to a slice, for example, woods[1:3] = ["Spruce", "Sugi",
"Rimu"]. The slice and the iterable don’t have to be the same length. In all cases,
the slice’s items are removed and the iterable’s items are inserted. This makes
the list shorter if the iterable has fewer items than the slice it replaces, and
longer if the iterable has more items than the slice.
To make what happens when assigning an iterable to a slice really clear, we
will consider one further example. Imagine that we have the list L = ["A", "B",
"C", "D", "E", "F"], and that we assign an iterable (in this case, a list) to a slice
of it with the code L[2:5] = ["X", "Y"]. First, the slice is removed, so behind the
scenes the list becomes ['A', 'B', 'F']. And then all the iterable’s items are
inserted at the slice’s start position, so the resultant list is ['A', 'B', 'X', 'Y',
Items can be removed in a number of other ways. We can use list.pop() with
no arguments to remove the rightmost item in a list—the removed item is also
returned. Similarly we can use list.pop() with an integer index argument to
remove (and return) an item at a particular index position. Another way of
removing an item is to call list.remove() with the item to be removed as the
argument. The del statement can also be used to remove individual items—for
example, del woods[4]—or to remove slices of items. Slices can also be removed
by assigning an empty list to a slice, so these two snippets are equivalent:
woods[2:4] = []
del woods[2:4]
In the left-hand snippet we have assigned an iterable (an empty list) to a
slice, so first the slice is removed, and since the iterable to insert is empty, no
insertion takes place.
69 ➤
When we first covered slicing and striding, we did so in the context of strings
where striding wasn’t very interesting. But in the case of lists, striding allows
us to access every n-th item which can often be useful. For example, suppose
we have the list, x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], and we want to set every
odd-indexed item (i.e., x[1], x[3], etc.) to 0. We can access every second item by
Chapter 3. Collection Data Types
striding, for example, x[::2]. But this will give us the items at index positions
0, 2, 4, and so on. We can fix this by giving an initial starting index, so now we
have x[1::2], and this gives us a slice of the items we want. To set each item
in the slice to 0, we need a list of 0s, and this list must have exactly the same
number of 0s as there are items in the slice.
Here is the complete solution: x[1::2] = [0] * len(x[1::2]). Now list x is [1,
0, 3, 0, 5, 0, 7, 0, 9, 0]. We used the replication operator *, to produce a list
consisting of the number of 0s we needed based on the length (i.e., the number
of items) of the slice. The interesting aspect is that when we assign the list [0,
0, 0, 0, 0] to the strided slice, Python correctly replaces x[1]’s value with the
first 0, x[3]’s value with the second 0, and so on.
Lists can be reversed and sorted in the same way as any other iterable using
the built-in reversed() and sorted() functions covered in the Iterators and Iterable Operations and Functions subsection (➤ 138). Lists also have equivalent
methods, list.reverse() and list.sort(), both of which work in-place (so they
don’t return anything), the latter accepting the same optional arguments as
sorted(). One common idiom is to case-insensitively sort a list of strings—for
example, we could sort the woods list like this: woods.sort(key=str.lower). The
key argument is used to specify a function which is applied to each item, and
whose return value is used to perform the comparisons used when sorting. As
we noted in the previous chapter’s section on string comparisons (68 ➤), for
languages other than English, sorting strings in a way that is meaningful to
humans can be quite challenging.
For inserting items, lists perform best when items are added or removed at the
end (list.append(), list.pop()). The worst performance occurs when we search
for items in a list, for example, using list.remove() or list.index(), or using in
for membership testing. If fast searching or membership testing is required,
a set or a dict (both covered later in this chapter) may be a more suitable
collection choice. Alternatively, lists can provide fast searching if they are kept
in order by sorting them—Python’s sort algorithm is especially well optimized
for sorting partially sorted lists—and using a binary search (provided by the
bisect module), to find items. (In Chapter 6 we will create an intrinsically
sorted custom list class.)
List Comprehensions
Small lists are often created using list literals, but longer lists are usually
created programmatically. For a list of integers we can use list(range(n)), or if
we just need an integer iterator, range() is sufficient, but for other lists using a
for … in loop is very common. Suppose, for example, that we wanted to produce
a list of the leap years in a given range. We might start out like this:
leaps = []
for year in range(1900, 1940):
➤ 140,
Sequence Types
if (year % 4 == 0 and year % 100 != 0) or (year % 400 == 0):
When the built-in range() function is given two integer arguments, n and m,
the iterator it returns produces the integers n, n + 1, …, m - 1.
Of course, if we knew the exact range beforehand we could use a list literal, for
example, leaps = [1904, 1908, 1912, 1916, 1920, 1924, 1928, 1932, 1936].
A list comprehension is an expression and a loop with an optional condition
enclosed in brackets where the loop is used to generate items for the list, and
where the condition can filter out unwanted items. The simplest form of a list
comprehension is this:
[item for item in iterable]
This will return a list of every item in the iterable, and is semantically no
different from list(iterable). Two things that make list comprehensions more
interesting and powerful are that we can use expressions, and we can attach a
condition—this takes us to the two general syntaxes for list comprehensions:
[expression for item in iterable]
[expression for item in iterable if condition]
The second syntax is equivalent to:
temp = []
for item in iterable:
if condition:
Normally, the expression will either be or involve the item. Of course, the
list comprehension does not need the temp variable needed by the for … in
loop version.
Now we can rewrite the code to generate the leaps list using a list comprehension. We will develop the code in three stages. First we will generate a list that
has all the years in the given range:
leaps = [y for y in range(1900, 1940)]
This could also be done using leaps = list(range(1900, 1940)). Now we’ll add a
simple condition to get every fourth year:
leaps = [y for y in range(1900, 1940) if y % 4 == 0]
Finally, we have the complete version:
leaps = [y for y in range(1900, 1940)
if (y % 4 == 0 and y % 100 != 0) or (y % 400 == 0)]
➤ 141
Chapter 3. Collection Data Types
Using a list comprehension in this case reduced the code from four lines to
two—a small savings, but one that can add up quite a lot in large projects.
Since list comprehensions produce lists, that is, iterables, and since the syntax
for list comprehensions requires an iterable, it is possible to nest list comprehensions. This is the equivalent of having nested for … in loops. For example,
if we wanted to generate all the possible clothing label codes for given sets of
sexes, sizes, and colors, but excluding labels for the full-figured females whom
the fashion industry routinely ignores, we could do so using nested for …
in loops:
codes = []
for sex in "MF":
for size in "SMLX":
if sex == "F" and size
for color in "BGW": #
codes.append(sex +
Male, Female
Small, Medium, Large, eXtra large
== "X":
Black, Gray, White
size + color)
This produces the 21 item list, ['MSB', 'MSG', …, 'FLW']. The same thing can be
achieved in just a couple of lines using a list comprehension:
codes = [s + z + c for s in "MF" for z in "SMLX" for c in "BGW"
if not (s == "F" and z == "X")]
Here, each item in the list is produced by the expression s + z + c. Also, we have
used subtly different logic for the list comprehension where we skip invalid
sex/size combinations in the innermost loop, whereas the nested for … in loops
version skips invalid combinations in its middle loop. Any list comprehension
can be rewritten using one or more for … in loops.
If the generated list is very large, it may be more efficient to generate each item
as it is needed rather than produce the whole list at once. This can be achieved
by using a generator rather than a list comprehension. We discuss this later,
in Chapter 8.
Set Types
A set type is a collection data type that supports the membership operator (in),
the size function (len()), and is iterable. In addition, set types at least provide
a set.isdisjoint() method, and support for comparisons, as well as support
for the bitwise operators (which in the context of sets are used for union,
intersection, etc.). Python provides two built-in set types: the mutable set type
and the immutable frozenset. When iterated, set types provide their items in
an arbitrary order.
➤ 341
Set Types
Only hashable objects may be added to a set. Hashable objects are objects
which have a __hash__() special method whose return value is always the same
throughout the object’s lifetime, and which can be compared for equality using
the __eq__() special method. (Special methods—methods whose name begins
and ends with two underscores—are covered in Chapter 6.)
All the built-in immutable data types, such as float, frozenset, int, str, and
tuple, are hashable and can be added to sets. The built-in mutable data types,
such as dict, list, and set, are not hashable since their hash value changes
depending on the items they contain, so they cannot be added to sets.
Set types can be compared using the standard comparison operators (<, <=, ==,
!=, >=, >). Note that although == and != have their usual meanings, with the
comparisons being applied item by item (and recursively for nested items such
as tuples or frozen sets inside sets), the other comparison operators perform
subset and superset comparisons, as we will see shortly.
A set is an unordered collection of zero or more object references that refer to
hashable objects. Sets are mutable, so we can easily add or remove items, but
since they are unordered they have no notion of index position and so cannot
be sliced or strided. Figure 3.3 illustrates the set created by the following
code snippet:
S = {7, "veil", 0, -29, ("x", 11), "sun", frozenset({8, 4, 7}), 913}
frozenset({8, 4, 7})
('x', 11)
Figure 3.3 A set is an unordered collection of unique items.
The set data type can be called as a function, set()—with no arguments it
returns an empty set, with a set argument it returns a shallow copy of the
argument, and with any other argument it attempts to convert the given object
to a set. It does not accept more than one argument. Nonempty sets can also
be created without using the set() function, but the empty set must be created
➤ 146
Set Types
Table 3.2 Set Methods and Operators
Adds item x to set s if it is not already in s
Removes all the items from set s
Returns a shallow copy of set s❄
s - t
Returns a new set that has every item that is in
set s that is not in set t❄
Removes every item that is in set t from set s
s -= t
Removes item x from set s if it is in s; see also
s & t
s &= t
s <= t
s >= t
Returns and removes a random item from set s,
or raises a KeyError exception if s is empty
Removes item x from set s, or raises a KeyError
exception if x is not in s; see also set.discard()
s ^ t
Returns a new set that has every item that is in
set s and every item that is in set t, but excluding items that are in both sets❄
s ^= t
s | t
Makes set s contain the symmetric difference of
itself and set t
s |= t
Returns a new set that has each item that is in
both set s and set t❄
Makes set s contain the intersection of itself
and set t
Returns True if sets s and t have no items in
Returns True if set s is equal to or a subset of set
t; use s < t to test whether s is a proper subset
of t❄
Returns True if set s is equal to or a superset
of set t; use s > t to test whether s is a proper
superset of t❄
Returns a new set that has all the items in set s
and all the items in set t that are not in set s❄
Adds every item in set t that is not in set s, to
set s
This method and its operator (if it has one) can also be used with frozensets.
➤ 146
Chapter 3. Collection Data Types
processing, once for each unique address. Assuming that the IP addresses are
hashable and are in iterable ips, and that the function we want called for each
one is called process_ip() and is already defined, the following code snippets
will do what we want, although with subtly different behavior:
seen = set()
for ip in ips:
if ip not in seen:
for ip in set(ips):
For the left-hand snippet, if we haven’t processed the IP address before, we add
it to the seen set and process it; otherwise, we ignore it. For the right-hand snippet, we only ever get each unique IP address to process in the first place. The
differences between the snippets are first that the left-hand snippet creates the
seen set which the right-hand snippet doesn’t need, and second that the lefthand snippet processes the IP addresses in the order they are encountered in
the ips iterable while the right-hand snippet processes them in an arbitrary
The right-hand approach is easier to code, but if the ordering of the ips
iterable is important we must either use the left-hand approach or change the
right-hand snippet’s first line to something like for ip in sorted(set(ips)): if
this is sufficient to get the required order. In theory the right-hand approach
might be slower if the number of items in ips is very large, since it creates the
set in one go rather than incrementally.
Sets are also used to eliminate unwanted items. For example, if we have a list
of filenames but don’t want any makefiles included (perhaps because they are
generated rather than handwritten), we might write:
filenames = set(filenames)
for makefile in {"MAKEFILE", "Makefile", "makefile"}:
This code will remove any makefile that is in the list using any of the standard
capitalizations. It will do nothing if no makefile is in the filenames list. The
same thing can be achieved in one line using the set difference (-) operator:
filenames = set(filenames) - {"MAKEFILE", "Makefile", "makefile"}
We can also use set.remove() to remove items, although this method raises a
KeyError exception if the item it is asked to remove is not in the set.
Set Types
Set Comprehensions
In addition to creating sets by calling set(), or by using a set literal, we can also
create sets using set comprehensions. A set comprehension is an expression and
a loop with an optional condition enclosed in braces. Like list comprehensions,
two syntaxes are supported:
{expression for item in iterable}
{expression for item in iterable if condition}
We can use these to achieve a filtering effect (providing the order doesn’t
matter). Here is an example:
html = {x for x in files if x.lower().endswith((".htm", ".html"))}
Given a list of filenames in files, this set comprehension makes the set html
hold only those filenames that end in .htm or .html, regardless of case.
Just like list comprehensions, the iterable used in a set comprehension can
itself be a set comprehension (or any other kind of comprehension), so quite
sophisticated set comprehensions can be created.
Frozen Sets
A frozen set is a set that, once created, cannot be changed. We can of course
rebind the variable that refers to a frozen set to refer to something else, though.
Frozen sets can only be created using the frozenset data type called as a
function. With no arguments, frozenset() returns an empty frozen set, with a
frozenset argument it returns a shallow copy of the argument, and with any
other argument it attempts to convert the given object to a frozenset. It does
not accept more than one argument.
Since frozen sets are immutable, they support only those methods and operators that produce a result without affecting the frozen set or sets to which
they are applied. Table 3.2 (123 ➤) lists all the set methods—frozen sets support frozenset.copy(), frozenset.difference() (-), frozenset.intersection() (&),
frozenset.isdisjoint(), frozenset.issubset() (<=; also < for proper subsets),
frozenset.issuperset() (>=; also > for proper supersets), frozenset.union() (|),
and frozenset.symmetric_difference() (^), all of which are indicated by a ❄ in
the table.
If a binary operator is used with a set and a frozen set, the data type of the
result is the same as the left-hand operand’s data type. So if f is a frozen set
and s is a set, f & s will produce a frozen set and s & f will produce a set. In the
case of the == and != operators, the order of the operands does not matter, and
f == s will produce True if both sets contain the same items.
➤ 146
Chapter 3. Collection Data Types
Another consequence of the immutability of frozen sets is that they meet
the hashable criterion for set items, so sets and frozen sets can contain frozen
We will see more examples of set use in the next section, and also in the
chapter’s Examples section.
Mapping Types
A mapping type is one that supports the membership operator (in) and the
size function (len()), and is iterable. Mappings are collections of key–value
items and provide methods for accessing items and their keys and values.
When iterated, unordered mapping types provide their items in an arbitrary
order. Python 3.0 provides two unordered mapping types, the built-in dict
type and the standard library’s collections.defaultdict type. A new, ordered
mapping type, collections.OrderedDict, was introduced with Python 3.1; this is
a dictionary that has the same methods and properties (i.e., the same API) as
the built-in dict, but stores its items in insertion order.★ We will use the term
dictionary to refer to any of these types when the difference doesn’t matter.
121 ➤
Only hashable objects may be used as dictionary keys, so immutable data types
such as float, frozenset, int, str, and tuple can be used as dictionary keys, but
mutable types such as dict, list, and set cannot. On the other hand, each key’s
associated value can be an object reference referring to an object of any type,
including numbers, strings, lists, sets, dictionaries, functions, and so on.
Dictionary types can be compared using the standard equality comparison operators (== and !=), with the comparisons being applied item by item (and recursively for nested items such as tuples or dictionaries inside dictionaries). Comparisons using the other comparison operators (<, <=, >=, >) are not supported
since they don’t make sense for unordered collections such as dictionaries.
A dict is an unordered collection of zero or more key–value pairs whose keys
are object references that refer to hashable objects, and whose values are object
references referring to objects of any type. Dictionaries are mutable, so we can
easily add or remove items, but since they are unordered they have no notion
of index position and so cannot be sliced or strided.
API stands for Application Programming Interface, a generic term used to refer to the public
methods and properties that classes provide, and to the parameters and return values of functions
and methods. For example, Python’s documentation documents the APIs that Python provides.
Mapping Types
The dict data type can be called as a function, dict()—with no arguments it
returns an empty dictionary, and with a mapping argument it returns a dictionary based on the argument; for example, returning a shallow copy if the
argument is a dictionary. It is also possible to use a sequence argument, providing that each item in the sequence is itself a sequence of two objects, the
first of which is used as a key and the second of which is used as a value.
Alternatively, for dictionaries where the keys are valid Python identifiers, keyword arguments can be used, with the key as the keyword and the value as the
key’s value. Dictionaries can also be created using braces—empty braces, {},
create an empty dictionary; nonempty braces must contain one or more commaseparated items, each of which consists of a key, a literal colon, and a value.
Another way of creating dictionaries is to use a dictionary comprehension—a
topic we will cover later in this subsection.
Here are some examples to illustrate the various syntaxes—they all produce
the same dictionary:
➤ 146
➤ 174
➤ 134
dict({"id": 1948, "name": "Washer", "size": 3})
dict(id=1948, name="Washer", size=3)
dict([("id", 1948), ("name", "Washer"), ("size", 3)])
dict(zip(("id", "name", "size"), (1948, "Washer", 3)))
{"id": 1948, "name": "Washer", "size": 3}
Dictionary d1 is created using a dictionary literal. Dictionary d2 is created using keyword arguments. Dictionaries d3 and d4 are created from sequences,
and dictionary d5 is created from a dictionary literal. The built-in zip() function that is used to create dictionary d4 returns a list of tuples, the first of which
has the first items of each of the zip() function’s iterable arguments, the second
of which has the second items, and so on. The keyword argument syntax (used
to create dictionary d2) is usually the most compact and convenient, providing
the keys are valid identifiers.
Figure 3.5 illustrates the dictionary created by the following code snippet:
d = {"root": 18, "blue": [75, "R", 2], 21: "venus", -14: None,
"mars": "rover", (4, 11): 18, 0: 45}
Dictionary keys are unique, so if we add a key–value item whose key is the
same as an existing key, the effect is to replace that key’s value with a new value. Brackets are used to access individual values—for example, d["root"] returns 18, d[21] returns the string "venus", and d[91] causes a KeyError exception
to be raised, given the dictionary shown in Figure 3.5.
Brackets can also be used to add and delete dictionary items. To add an item
we use the = operator, for example, d["X"] = 59. And to delete an item we use
the del statement—for example, del d["mars"] will delete the item whose key
is “mars” from the dictionary, or raise a KeyError exception if no item has that
➤ 143
Chapter 3. Collection Data Types
(4, 11)
[75, 'R', 2]
Figure 3.5 A dictionary is an unsorted collection of (key, value) items with unique keys.
key. Items can also be removed (and returned) from the dictionary using the
dict.pop() method.
Dictionaries support the built-in len() function, and for their keys, fast
membership testing with in and not in. All the dictionary methods are listed in
Table 3.3.
Because dictionaries have both keys and values, we might want to iterate over
a dictionary by (key, value) items, by values, or by keys. For example, here are
two equivalent approaches to iterating by (key, value) pairs:
for item in d.items():
print(item[0], item[1])
for key, value in d.items():
print(key, value)
Iterating over a dictionary’s values is very similar:
for value in d.values():
To iterate over a dictionary’s keys we can use dict.keys(), or we can simply
treat the dictionary as an iterable that iterates over its keys, as these two
equivalent code snippets illustrate:
for key in d:
for key in d.keys():
If we want to change the values in a dictionary, the idiom to use is to iterate
over the keys and change the values using the brackets operator. For example,
here is how we would increment every value in dictionary d, assuming that all
the values are numbers:
for key in d:
d[key] += 1
Mapping Types
Table 3.3 Dictionary Methods
Removes all items from dict d
Returns a shallow copy of dict d
s, v)
Returns a dict whose keys are the items in sequence s and
whose values are None or v if v is given
Returns key k’s associated value, or None if k isn’t in dict d
➤ 146
d.get(k, v)
Returns key k’s associated value, or v if k isn’t in dict d
Returns a view★ of all the (key, value) pairs in dict d
Returns a view★ of all the keys in dict d
Returns key k’s associated value and removes the item
whose key is k, or raises a KeyError exception if k isn’t in d
d.pop(k, v)
Returns key k’s associated value and removes the item
whose key is k, or returns v if k isn’t in dict d
Returns and removes an arbitrary (key, value) pair from
dict d, or raises a KeyError exception if d is empty
k, v)
The same as the dict.get() method, except that if the key is
not in dict d, a new item is inserted with the key k, and with
a value of None or of v if v is given
Adds every (key, value) pair from a that isn’t in dict d to d,
and for every key that is in both d and a, replaces the corresponding value in d with the one in a—a can be a dictionary,
an iterable of (key, value) pairs, or keyword arguments
Returns a view★ of all the values in dict d
The dict.items(), dict.keys(), and dict.values() methods all return dictionary
views. A dictionary view is effectively a read-only iterable object that appears
to hold the dictionary’s items or keys or values, depending on the view we have
asked for.
In general, we can simply treat views as iterables. However, two things make
a view different from a normal iterable. One is that if the dictionary the view
refers to is changed, the view reflects the change. The other is that key and
item views support some set-like operations. Given dictionary view v and set
or dictionary view x, the supported operations are:
v & x
v | x
# Intersection
# Union
Dictionary views can be thought of—and used as—iterables; they are discussed in the text.
Chapter 3. Collection Data Types
v - x
v ^ x
# Difference
# Symmetric difference
We can use the membership operator, in, to see whether a particular key is in
a dictionary, for example, x in d. And we can use the intersection operator to see
which keys from a given set are in a dictionary. For example:
d = {}.fromkeys("ABCD", 3)
s = set("ACX")
matches = d.keys() & s
# d == {'A': 3, 'B': 3, 'C': 3, 'D': 3}
# s == {'A', 'C', 'X'}
# matches == {'A', 'C'}
Note that in the snippet’s comments we have used alphabetical order—this is
purely for ease of reading since dictionaries and sets are unordered.
Dictionaries are often used to keep counts of unique items. One such example
of this is counting the number of occurrences of each unique word in a file.
Here is a complete program ( that lists every word and the
number of times it occurs in alphabetical order for all the files listed on the
command line:
import string
import sys
words = {}
strip = string.whitespace + string.punctuation + string.digits + "\"'"
for filename in sys.argv[1:]:
for line in open(filename):
for word in line.lower().split():
word = word.strip(strip)
if len(word) > 2:
words[word] = words.get(word, 0) + 1
for word in sorted(words):
print("'{0}' occurs {1} times".format(word, words[word]))
We begin by creating an empty dictionary called words. Then we create a string
that contains all those characters that we want to ignore, by concatenating
some useful strings provided by the string module. We iterate over each filename given on the command line, and over each line in each file. See the sidebar “Reading and Writing Text Files” (➤ 131) for an explanation of the open()
function. We don’t specify an encoding (because we don’t know what each file’s
encoding will be), so we let Python open each file using the default local encoding. We split each lowercased line into words, and then strip off the characters
that we want to ignore from both ends of each word. If the resultant word is
at least three characters long we need to update the dictionary.
We cannot use the syntax words[word] += 1 because this will raise a KeyError
exception the first time a new word is encountered—after all, we can’t increment the value of an item that does not yet exist in the dictionary. So we use
Mapping Types
Reading and Writing Text Files
Files are opened using the built-in open() function, which returns a “file
object” (of type io.TextIOWrapper for text files). The open() function takes one
mandatory argument—the filename, which may include a path—and up
to six optional arguments, two of which we briefly cover here. The second
argument is the mode—this is used to specify whether the file is to be treated
as a text file or as a binary file, and whether the file is to be opened for
reading, writing, appending, or a combination of these.
91 ➤
For text files, Python uses an encoding that is platform-dependent. Where
possible it is best to specify the encoding using open()’s encoding argument,
so the syntaxes we normally use for opening files are these:
fin = open(filename, encoding="utf8")
fout = open(filename, "w", encoding="utf8")
# for reading text
# for writing text
Because open()’s mode defaults to “read text”, and by using a keyword rather
than a positional argument for the encoding argument, we can omit the other
optional positional arguments when opening for reading. And similarly,
when opening to write we need to give only the arguments we actually want
to use. (Argument passing is covered in depth in Chapter 4.)
Once a file is opened for reading in text mode, we can read the whole file into
a single string using the file object’s read() method, or into a list of strings
using the file object’s readlines() method. A very common idiom for reading
line by line is to treat the file object as an iterator:
for line in open(filename, encoding="utf8"):
This works because a file object can be iterated over, just like a sequence,
with each successive item being a string containing the next line from the
file. The lines we get back include the line termination character, \n.
If we specify a mode of “w”, the file is opened in “write text” mode. We write
to a file using the file object’s write() method, which takes a single string as
its argument. Each line written should end with a \n. Python automatically
translates between \n and the underlying platform’s line termination
characters when reading and writing.
Once we have finished using a file object we can call its close() method—this
will cause any outstanding writes to be flushed. In small Python programs
it is very common not to bother calling close(), since Python does this
automatically when the file object goes out of scope. If a problem occurs, it
will be indicated by an exception being raised.
Chapter 7
➤ 287
Chapter 3. Collection Data Types
a subtler approach. We call dict.get() with a default value of 0. If the word
is already in the dictionary, dict.get() will return the associated number, and
this value plus 1 will be set as the item’s new value. If the word is not in the
dictionary, dict.get() will return the supplied default of 0, and this value plus
1 (i.e., 1) will be set as the value of a new item whose key is the string held by
word. To clarify, here are two code snippets that do the same thing, although the
code using dict.get() is more efficient:
words[word] = words.get(word, 0) + 1
if word not in words:
words[word] = 0
words[word] += 1
In the next subsection where we cover default dictionaries, we will see an
alternative solution.
Once we have accumulated the dictionary of words, we iterate over its keys
(the words) in sorted order, and print each word and the number of times
it occurs.
Using dict.get() allows us to easily update dictionary values, providing the
values are single items like numbers or strings. But what if each value is itself
a collection? To demonstrate how to handle this we will look at a program
that reads HTML files given on the command line and prints a list of each
unique Web site that is referred to in the files with a list of the referring files
listed indented below the name of each Web site. Structurally, the program
( is very similar to the unique words program we have just
reviewed. Here is the main part of the code:
sites = {}
for filename in sys.argv[1:]:
for line in open(filename):
i = 0
while True:
site = None
i = line.find("http://", i)
if i > -1:
i += len("http://")
for j in range(i, len(line)):
if not (line[j].isalnum() or line[j] in ".-"):
site = line[i:j].lower()
if site and "." in site:
sites.setdefault(site, set()).add(filename)
i = j
Mapping Types
We begin by creating an empty dictionary. Then we iterate over each file listed
on the command line and each line within each file. We must account for the
fact that each line may refer to any number of Web sites, which is why we keep
calling str.find() until it fails. If we find the string “http://”, we increment i
(our starting index position) by the length of “http://”, and then we look at each
succeeding character until we reach one that isn’t valid for a Web site’s name.
If we find a site (and as a simply sanity check, only if it contains a period), we
add it to the dictionary.
We cannot use the syntax sites[site].add(filename) because this will raise a
KeyError exception the first time a new site is encountered—after all, we can’t
add to a set that is the value of an item that does not yet exist in the dictionary.
So we must use a different approach. The dict.setdefault() method returns an
object reference to the item in the dictionary that has the given key (the first
argument). If there is no such item, the method creates a new item with the
key and sets its value either to None, or to the given default value (the second
argument). In this case we pass a default value of set(), that is, an empty set.
So the call to dict.setdefault() always returns an object reference to a value,
either one that existed before or a new one. (Of course, if the given key is not
hashable a TypeError exception will be raised.)
In this example, the returned object reference always refers to a set (an empty
set the first time any particular key, that is, site, is encountered), and we then
add the filename that refers to the site to the site’s set of filenames. By using
a set we ensure that even if a file refers to a site repeatedly, we record the
filename only once for the site.
To make the dict.setdefault() method’s functionality clear, here are two
equivalent code snippets:
sites.setdefault(site, set()).add(fname)
if site not in sites:
sites[site] = set()
For the sake of completeness, here is the rest of the program:
for site in sorted(sites):
print("{0} is referred to in:".format(site))
for filename in sorted(sites[site], key=str.lower):
Each Web site is printed with the files that refer to it printed indented underneath. The sorted() call in the outer for … in loop sorts all the dictionary’s
keys—whenever a dictionary is used in a context that requires an iterable it is
the keys that are used. If we want the iterable to be the (key, value) items or
the values, we can use dict.items() or dict.values(). The inner for … in loop
iterates over the sorted filenames from the current site’s set of filenames.
➤ 140,
Using str.
mapping unpacking
81 ➤
Chapter 3. Collection Data Types
Although a dictionary of web sites is likely to contain a lot of items, many
other dictionaries have only a few items. For small dictionaries, we can print
their contents using their keys as field names and using mapping unpacking
to convert the dictionary’s key–value items into key–value arguments for the
str.format() method.
➤ 177
>>> greens = dict(green="#0080000", olive="#808000", lime="#00FF00")
>>> print("{green} {olive} {lime}".format(**greens))
#0080000 #808000 #00FF00
Here, using mapping unpacking (**) has exactly the same effect as writing
.format(, olive=greens.olive, lime=greens.lime), but is easier to write and arguably clearer. Note that it doesn’t matter if the dictionary
has more keys than we need, since only those keys whose names appear in the
format string are used.
Dictionary Comprehensions
A dictionary comprehension is an expression and a loop with an optional
condition enclosed in braces, very similar to a set comprehension. Like list and
set comprehensions, two syntaxes are supported:
{keyexpression: valueexpression for key, value in iterable}
{keyexpression: valueexpression for key, value in iterable if condition}
Here is how we could use a dictionary comprehension to create a dictionary
where each key is the name of a file in the current directory and each value is
the size of the file in bytes:
file_sizes = {name: os.path.getsize(name) for name in os.listdir(".")}
The os (“operating system”) module’s os.listdir() function returns a list of
the files and directories in the path it is passed, although it never includes
“.” or “..” in the list. The os.path.getsize() function returns the size of the
given file in bytes. We can avoid directories and other nonfile entries by adding
a condition:
file_sizes = {name: os.path.getsize(name) for name in os.listdir(".")
if os.path.isfile(name)}
The os.path module’s os.path.isfile() function returns True if the path passed
to it is that of a file, and False otherwise—that is, for directories, links, and
so on.
A dictionary comprehension can also be used to create an inverted dictionary.
For example, given dictionary d, we can produce a new dictionary whose keys
are d’s values and whose values are d’s keys:
os and
➤ 224
Mapping Types
inverted_d = {v: k for k, v in d.items()}
The resultant dictionary can be inverted back to the original dictionary if all
the original dictionary’s values are unique—but the inversion will fail with a
TypeError being raised if any value is not hashable.
Just like list and set comprehensions, the iterable in a dictionary comprehension can be another comprehension, so all kinds of nested comprehensions are
Default Dictionaries
Default dictionaries are dictionaries—they have all the operators and methods
that dictionaries provide. What makes default dictionaries different from
plain dictionaries is the way they handle missing keys; in all other respects
they behave identically to dictionaries. (In object-oriented terms, defaultdict
is a subclass of dict; object-oriented programming, including subclassing, is
covered in Chapter 6.)
If we use a nonexistent (“missing”) key when accessing a dictionary, a KeyError
is raised. This is useful because we often want to know whether a key that we
expected to be present is absent. But in some cases we want every key we use
to be present, even if it means that an item with the key is inserted into the
dictionary at the time we first access it.
For example, if we have a dictionary d which does not have an item with
key m, the code x = d[m] will raise a KeyError exception. But if d is a suitably
created default dictionary, if an item with key m is in the default dictionary, the
corresponding value is returned the same as for a dictionary—but if m is not a
key in the default dictionary, a new item with key m is created with a default
value, and the newly created item’s value is returned.
130 ➤
Earlier we wrote a small program that counted the unique words in the
files it was given on the command line. The dictionary of words was created
like this:
words = {}
Each key in the words dictionary was a word and each value an integer holding
the number of times the word had occurred in all the files that were read.
Here’s how we incremented whenever a suitable word was encountered:
words[word] = words.get(word, 0) + 1
We had to use dict.get() to account for when the word was encountered the
first time (where we needed to create a new item with a count of 1) and for
when the word was encountered subsequently (where we needed to add 1 to the
word’s existing count).
Chapter 3. Collection Data Types
When a default dictionary is created, we can pass in a factory function. A factory
function is a function that, when called, returns an object of a particular type.
All of Python’s built-in data types can be used as factory functions, for example,
data type str can be called as str()—and with no argument it returns an empty string object. The factory function passed to a default dictionary is used to
create default values for missing keys.
Note that the name of a function is an object reference to the function—so
when we want to pass functions as parameters, we just pass the name. When
we use a function with parentheses, the parentheses tell Python that the
function should be called.
The program has one more line than the original program (import collections), and the lines for creating and updating
the dictionary are written differently. Here is how the default dictionary is
words = collections.defaultdict(int)
The words default dictionary will never raise a KeyError. If we were to write
x = words["xyz"] and there was no item with key "xyz", when the access is
attempted and the key isn’t found, the default dictionary will immediately
create a new item with key "xyz" and value 0 (by calling int()), and this value
is what will be assigned to x.
words[word] += 1
Now we no longer need to use dict.get(); instead we can simply increment the
item’s value. The very first time a word is encountered, a new item is created
with value 0 (to which 1 is immediately added), and on every subsequent
access, 1 is added to whatever the current value happens to be.
We have now completed our review of all of Python’s built-in collection data
types, and a couple of the standard library’s collection data types. In the next
section we will look at some issues that are common to all of the collection data
Ordered Dictionaries
The ordered dictionaries type—collections.OrderedDict—was introduced with
Python 3.1 in fulfillment of PEP 372. Ordered dictionaries can be used as
drop-in replacements for unordered dicts because they provide the same API.
The difference between the two is that ordered dictionaries store their items in
the order in which they were inserted—a feature that can be very convenient.
Note that if an ordered dictionary is passed an unordered dict or keyword arguments when it is created, the item order will be arbitrary; this is because under the hood Python passes keyword arguments using a standard unordered
Mapping Types
dict. A similar effect occurs with the use of the update() method. For these
reasons, passing keyword arguments or an unordered dict when creating an
ordered dictionary or using update() on one is best avoided. However, if we pass
a list or tuple of key–value 2-tuples when creating an ordered dictionary, the
ordering is preserved (since they are passed as a single item—a list or tuple).
Here’s how to create an ordered dictionary using a list of 2-tuples:
d = collections.OrderedDict([('z', -4), ('e', 19), ('k', 7)])
Because we used a single list as argument the key ordering is preserved. It is
probably more common to create ordered dictionaries incrementally, like this:
tasks = collections.OrderedDict()
tasks[8031] = "Backup"
tasks[4027] = "Scan Email"
tasks[5733] = "Build System"
If we had created unordered dicts the same way and asked for their keys, the
order of the returned keys would be arbitrary. But for ordered dictionaries, we
can rely on the keys to be returned in the same order they were inserted. So
for these examples, if we wrote list(d.keys()), we are guaranteed to get the list
['z', 'e', 'k'], and if we wrote list(tasks.keys()), we are guaranteed to get
the list [8031, 4027, 5733].
One other nice feature of ordered dictionaries is that if we change an item’s
value—that is, if we insert an item with the same key as an existing key—the
order is not changed. So if we did tasks[8031] = "Daily backup", and then asked
for the list of keys, we would get exactly the same list in exactly the same order
as before.
If we want to move an item to the end, we must delete it and then reinsert it.
We can also call popitem() to remove and return the last key–value item in the
ordered dictionary; or we can call popitem(last=False), in which case the first
item will be removed and returned.
Another, slightly more specialized use for ordered dictionaries is to produce
sorted dictionaries. Given a dictionary, d, we can convert it into a sorted
dictionary like this: d = collections.OrderedDict(sorted(d.items())). Note that
if we were to insert any additional keys they would be inserted at the end, so
after any insertion, to preserve the sorted order, we would have to re-create the
dictionary by executing the same code we used to create it in the first place.
Doing insertions and re-creating isn’t quite as inefficient as it sounds, since
Python’s sorting algorithm is highly optimized, especially for partially sorted
data, but it is still potentially expensive.
In general, using an ordered dictionary to produce a sorted dictionary makes
sense only if we expect to iterate over the dictionary multiple times, and if we
do not expect to do any insertions (or very few), once the sorted dictionary has
Chapter 3. Collection Data Types
been created. (An implementation of a real sorted dictionary that automatically maintains its keys in sorted order is presented in Chapter 6; ➤ 276.)
Iterating and Copying Collections
Once we have collections of data items, it is natural to want to iterate over all
the items they contain. In this section’s first subsection we will introduce some
of Python’s iterators and the operators and functions that involve iterators.
Another common requirement is to copy a collection. There are some subtleties
involved here because of Python’s use of object references (for the sake of
efficiency), so in this section’s second subsection, we will examine how to copy
collections and get the behavior we want.
Iterators and Iterable Operations and Functions
An iterable data type is one that can return each of its items one at a time. Any
object that has an __iter__() method, or any sequence (i.e., an object that has a
__getitem__() method taking integer arguments starting from 0) is an iterable
and can provide an iterator. An iterator is an object that provides a __next__()
method which returns each successive item in turn, and raises a StopIteration
exception when there are no more items. Table 3.4 lists the operators and
functions that can be used with iterables.
The order in which items are returned depends on the underlying iterable. In
the case of lists and tuples, items are normally returned in sequential order
starting from the first item (index position 0), but some iterators return the
items in an arbitrary order—for example, dictionary and set iterators.
The built-in iter() function has two quite different behaviors. When given
a collection data type or a sequence it returns an iterator for the object it is
passed—or raises a TypeError if the object cannot be iterated. This use arises
when creating custom collection data types, but is rarely needed in other contexts. The second iter() behavior occurs when the function is passed a callable
(a function or method), and a sentinel value. In this case the function passed in
is called once at each iteration, returning the function’s return value each time,
or raising a StopIteration exception if the return value equals the sentinel.
When we use a for item in iterable loop, Python in effect calls iter(iterable)
to get an iterator. This iterator’s __next__() method is then called at each loop
iteration to get the next item, and when the StopIteration exception is raised,
it is caught and the loop is terminated. Another way to get an iterator’s next
item is to call the built-in next() function. Here are two equivalent pieces of
code (multiplying the values in a list), one using a for … in loop and the other
using an explicit iterator:
➤ 274
Iterating and Copying Collections
product = 1
for i in [1, 2, 4, 8]:
product *= i
print(product) # prints: 64
product = 1
i = iter([1, 2, 4, 8])
while True:
product *= next(i)
except StopIteration:
print(product) # prints: 64
Any (finite) iterable, i, can be converted into a tuple by calling tuple(i), or can
be converted into a list by calling list(i).
The all() and any() functions can be used on iterators and are often used in
functional-style programming. Here are a couple of usage examples that show
all(), any(), len(), min(), max(), and sum():
>>> x = [-2, 9, 7, -4, 3]
>>> all(x), any(x), len(x), min(x), max(x), sum(x)
(True, True, 5, -4, 9, 13)
>>> x.append(0)
>>> all(x), any(x), len(x), min(x), max(x), sum(x)
(False, True, 6, -4, 9, 13)
Of these little functions, len() is probably the most frequently used.
The enumerate() function takes an iterator and returns an enumerator object.
This object can be treated like an iterator, and at each iteration it returns a
2-tuple with the tuple’s first item the iteration number (by default starting
from 0), and the second item the next item from the iterator enumerate() was
called on. Let’s look at enumerate()’s use in the context of a tiny but complete
The program takes a word and one or more filenames on the
command line and outputs the filename, line number, and line whenever the
line contains the given word.★ Here’s a sample run: Dom data/forenames.txt
Data files data/forenames.txt and data/surnames.txt contain unsorted lists of
names, one per line.
In Chapter 10 will see two other implementations of this program, and, which spread the work over multiple processes and multiple threads.
➤ 395
Chapter 3. Collection Data Types
Table 3.4 Common Iterable Operators and Functions
s + t
Returns a sequence that is the concatenation of sequences s
and t
Returns a sequence that is int n concatenations of sequence s
s * n
x in i
Returns True if item x is in iterable i; use not in to reverse
the test
Returns True if every item in iterable i evaluates to True
Returns True if any item in iterable i evaluates to True
enumerate(i, Normally used in for … in loops to provide a sequence of (instart)
dex, item) tuples with indexes starting at 0 or start; see text
Returns the “length” of x. If x is a collection it is the number
of items; if x is a string it is the number of characters.
max(i, key)
Returns the biggest item in iterable i or the item with the
biggest key(item) value if a key function is given
min(i, key)
Returns the smallest item in iterable i or the item with the
smallest key(item) value if a key function is given
Returns an integer iterator. With one argument (stop), the iterator goes from 0 to stop - 1; with two arguments (start, stop)
the iterator goes from start to stop - 1; with three arguments
it goes from start to stop - 1 in steps of step.
Returns an iterator that returns the items from iterator i in
reverse order
Returns a list of the items from iterator i in sorted order; key
is used to provide DSU (Decorate, Sort, Undecorate) sorting.
reverse) If reverse is True the sorting is done in reverse order.
Returns the sum of the items in iterable i plus start (which
defaults to 0); i may not contain strings
Returns an iterator of tuples using the iterators i1 to iN;
..., iN) see text
Apart from the sys import, the program is just ten lines long:
if len(sys.argv) < 3:
print("usage: word infile1 [infile2 [... infileN]]")
word = sys.argv[1]
for filename in sys.argv[2:]:
for lino, line in enumerate(open(filename), start=1):
if word in line:
Iterating and Copying Collections
print("{0}:{1}:{2:.40}".format(filename, lino,
We begin by checking that there are at least two command-line arguments.
If there are not, we print a usage message and terminate the program. The
sys.exit() function performs an immediate clean termination, closing any open
files. It accepts an optional int argument which is passed to the calling shell.
Reading and
text files
131 ➤
We assume that the first argument is the word the user is looking for and that
the other arguments are the names of the files to look in. We have deliberately
called open() without specifying an encoding—the user might use wildcards
to specify any number of files, each potentially with a different encoding, so in
this case we leave Python to use the platform-dependent encoding.
The file object returned by the open() function in text mode can be used as an
iterator, returning one line of the file on each iteration. By passing the iterator to enumerate(), we get an enumerator iterator that returns the iteration
number (in variable lino, “line number”) and a line from the file, on each iteration. If the word the user is looking for is in the line, we print the filename, line
number, and the first 40 characters of the line with trailing whitespace (e.g.,
\n) stripped. The enumerate() function accepts an optional keyword argument,
start, which defaults to 0; we have used this argument set to 1, since by convention, text file line numbers are counted from 1.
Quite often we don’t need an enumerator, but rather an iterator that returns
successive integers. This is exactly what the range() function provides. If we
need a list or tuple of integers, we can convert the iterator returned by range()
by using the appropriate conversion function. Here are a few examples:
>>> list(range(5)), list(range(9, 14)), tuple(range(10, -11, -5))
([0, 1, 2, 3, 4], [9, 10, 11, 12, 13], (10, 5, 0, -5, -10))
The range() function is most commonly used for two purposes: to create lists or
tuples of integers, and to provide loop counting in for … in loops. For example,
these two equivalent examples ensure that list x’s items are all non-negative:
for i in range(len(x)):
x[i] = abs(x[i])
i = 0
while i < len(x):
x[i] = abs(x[i])
i += 1
In both cases, if list x was originally, say, [11, -3, -12, 8, -1], afterward it will
be [11, 3, 12, 8, 1].
Since we can unpack an iterable using the * operator, we can unpack the
iterator returned by the range() function. For example, if we have a function
called calculate() that takes four arguments, here are some ways we could call
it with arguments, 1, 2, 3, and 4:
Chapter 3. Collection Data Types
calculate(1, 2, 3, 4)
t = (1, 2, 3, 4)
calculate(*range(1, 5))
In all three calls, four arguments are passed. The second call unpacks a 4-tuple,
and the third call unpacks the iterator returned by the range() function.
We will now look at a small but complete program to consolidate some of the
things we have covered so far, and for the first time to explicitly write to a file.
The program reads in a file of forenames and a file
of surnames, creating two lists, and then creates the file test-names1.txt and
writes 100 random names into it.
We will use the random.choice() function which takes a random item from a
sequence, so it is possible that some duplicate names might occur. First we’ll
look at the function that returns the lists of names, and then we will look at
the rest of the program.
def get_forenames_and_surnames():
forenames = []
surnames = []
for names, filename in ((forenames, "data/forenames.txt"),
(surnames, "data/surnames.txt")):
for name in open(filename, encoding="utf8"):
return forenames, surnames
110 ➤
In the outer for … in loop, we iterate over two 2-tuples, unpacking each 2-tuple
into two variables. Even though the two lists might be quite large, returning
them from a function is efficient because Python uses object references, so the
only thing that is really returned is a tuple of two object references.
Inside Python programs it is convenient to always use Unix-style paths, since
they can be typed without the need for escaping, and they work on all platforms
(including Windows). If we have a path we want to present to the user in, say,
variable path, we can always import the os module and call path.replace("/",
os.sep) to replace forward slashes with the platform-specific directory separator.
forenames, surnames = get_forenames_and_surnames()
fh = open("test-names1.txt", "w", encoding="utf8")
for i in range(100):
line = "{0} {1}\n".format(random.choice(forenames),
Iterating and Copying Collections
Reading and
text files
131 ➤
Having retrieved the two lists we open the output file for writing, and keep
the file object in variable fh (“file handle”). We then loop 100 times, and in each
iteration we create a line to be written to the file, remembering to include a
newline at the end of every line. We make no use of the loop variable i; it is
needed purely to satisfy the for … in loop’s syntax. The preceding code snippet,
the get_forenames_and_surnames() function, and an import statement constitute
the entire program.
In the program we paired items from two separate
lists together into strings. Another way of combining items from two or
more lists (or other iterables) is to use the zip() function. The zip() function
takes one or more iterables and returns an iterator that returns tuples. The
first tuple has the first item from every iterable, the second tuple the second
item from every iterable, and so on, stopping as soon as one of the iterables is
exhausted. Here is an example:
for t in zip(range(4), range(0, 10, 2), range(1, 10, 2)):
0, 1)
2, 3)
4, 5)
6, 7)
Although the iterators returned by the second and third range() calls can
produce five items each, the first can produce only four, so that limits the
number of items zip() can return to four tuples.
Here is a modified version of the program to generate test names, this time
with each name occupying 25 characters and followed by a random year. The
program is called and outputs the file test-names2.txt.
We have not shown the get_forenames_and_surnames() function or the open() call
since, apart from the output filename, they are the same as before.
limit = 100
years = list(range(1970, 2013)) * 3
for year, forename, surname in zip(
random.sample(years, limit),
random.sample(forenames, limit),
random.sample(surnames, limit)):
name = "{0} {1}".format(forename, surname)
fh.write("{0:.<25}.{1}\n".format(name, year))
We begin by setting a limit on how many names we want to generate. Then we
create a list of years by making a list of the years from 1970 to 2012 inclusive,
and then replicating this list three times so that the final list has three occurrences of each year. This is necessary because the random.sample() function
that we are using (instead of random.choice()) takes both an iterable and how
Chapter 3. Collection Data Types
many items it is to produce—a number that cannot be less than the number
of items the iterable can return. The random.sample() function returns an iterator that will produce up to the specified number of items from the iterable it
is given—with no repeats. So this version of the program will always produce
unique names.
110 ➤
78 ➤
In the for … in loop we unpack each tuple returned by the zip() function. We
want to limit the length of each name to 25 characters, and to do this we must
first create a string with the complete name, and then set the maximum width
for that string when we call str.format() the second time. We left-align each
name, and for names shorter than 25 characters we fill with periods. The extra
period ensures that names that occupy the full field width are still separated
from the year by a period.
We will conclude this subsection by mentioning two other iterable-related
functions, sorted() and reversed(). The sorted() function returns a list with
the items sorted, and the reversed() function simply returns an iterator that
iterates in the reverse order to the iterator it is given as its argument. Here is
an example of reversed():
1, 2, 3, 4, 5]
4, 3, 2, 1, 0]
The sorted() function is more sophisticated, as these examples show:
>>> x = []
>>> for t in zip(range(-10, 0, 1), range(0, 10, 2), range(1, 10, 2)):
x += t
>>> x
[-10, 0, 1, -9, 2, 3, -8, 4, 5, -7, 6, 7, -6, 8, 9]
>>> sorted(x)
[-10, -9, -8, -7, -6, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
>>> sorted(x, reverse=True)
[9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -6, -7, -8, -9, -10]
>>> sorted(x, key=abs)
[0, 1, 2, 3, 4, 5, 6, -6, -7, 7, -8, 8, -9, 9, -10]
In the preceding snippet, the zip() function returns 3-tuples, (-10, 0, 1), (-9,
2, 3), and so on. The += operator extends a list, that is, it appends each item in
the sequence it is given to the list.
The first call to sorted() returns a copy of the list using the conventional sort
order. The second call returns a copy of the list in the reverse of the conventional sort order. The last call to sorted() specifies a “key” function which we
will come back to in a moment.
Iterating and Copying Collections
Notice that since Python functions are objects like any other, they can be
passed as arguments to other functions, and stored in collections without
formality. Recall that a function’s name is an object reference to the function; it
is the parentheses that follow the name that tell Python to call the function.
When a key function is passed (in this case the abs() function), it is called
once for every item in the list (with the item passed as the function’s sole
parameter), to create a “decorated” list. Then the decorated list is sorted, and
the sorted list without the decoration is returned as the result. We are free to
use our own custom function as the key function, as we will see shortly.
For example, we can case-insensitively sort a list of strings by passing the
str.lower() method as a key. If we have the list, x, of ["Sloop", "Yawl",
"Cutter", "schooner", "ketch"], we can sort it case-insensitively using DSU
(Decorate, Sort, Undecorate) with a single line of code by passing a key function, or do the DSU explicitly, as these two equivalent code snippets show:
x = sorted(x, key=str.lower)
temp = []
for item in x:
temp.append((item.lower(), item))
x = []
for key, value in sorted(temp):
Both snippets produce a new list: ["Cutter", "ketch", "schooner", "Sloop",
"Yawl"], although the computations they perform are not identical because the
right-hand snippet creates the temp list variable.
Python’s sort algorithm is an adaptive stable mergesort that is both fast and
smart, and it is especially well optimized for partially sorted lists—a very
common case.★ The “adaptive” part means that the sort algorithm adapts to
circumstances—for example, taking advantage of partially sorted data. The
“stable” part means that items that sort equally are not moved in relation to
each other (after all, there is no need), and the “mergesort” part is the generic
name for the sorting algorithm used. When sorting collections of integers,
strings, or other simple types their “less than” operator (<) is used. Python
can sort collections that contain collections, working recursively to any depth.
For example:
>>> x = list(zip((1, 3, 1, 3), ("pram", "dorie", "kayak", "canoe")))
>>> x
[(1, 'pram'), (3, 'dorie'), (1, 'kayak'), (3, 'canoe')]
>>> sorted(x)
[(1, 'kayak'), (1, 'pram'), (3, 'canoe'), (3, 'dorie')]
The algorithm was created by Tim Peters. An interesting explanation and discussion of the
algorithm is in the file listsort.txt which comes with Python’s source code.
Chapter 3. Collection Data Types
Python has sorted the list of tuples by comparing the first item of each tuple,
and when these are the same, by comparing the second item. This gives a
sort order based on the integers, with the strings being tiebreakers. We can
force the sort to be based on the strings and use the integers as tiebreakers by
defining a simple key function:
def swap(t):
return t[1], t[0]
The swap() function takes a 2-tuple and returns a new 2-tuple with the arguments swapped. Assuming that we have entered the swap() function in IDLE,
we can now do this:
>>> sorted(x, key=swap)
[(3, 'canoe'), (3, 'dorie'), (1, 'kayak'), (1, 'pram')]
Lists can also be sorted in-place using the list.sort() method, which takes the
same optional arguments as sorted().
Sorting can be applied only to collections where all the items can be compared
with each other:
sorted([3, 8, -7.5, 0, 1.3])
sorted([3, "spanner", -7.5, 0, 1.3])
# returns: [-7.5, 0, 1.3, 3, 8]
# raises a TypeError
Although the first list has numbers of different types (int and float), these
types can be compared with each other so that sorting a list containing them
works fine. But the second list has a string and this cannot be sensibly compared with a number, and so a TypeError exception is raised. If we want to sort
a list that has integers, floating-point numbers, and strings that contain numbers, we can give float() as the key function:
sorted(["1.3", -7.5, "5", 4, "-2.4", 1], key=float)
This returns the list [-7.5, '-2.4', 1, '1.3', 4, '5']. Notice that the list’s values
are not changed, so strings remain strings. If any of the strings cannot be
converted to a number (e.g., “spanner”), a ValueError exception will be raised.
Copying Collections
16 ➤
Since Python uses object references, when we use the assignment operator (=),
no copying takes place. If the right-hand operand is a literal such as a string
or a number, the left-hand operand is set to be an object reference that refers to
the in-memory object that holds the literal’s value. If the right-hand operand
is an object reference, the left-hand operand is set to be an object reference that
refers to the same object as the right-hand operand. One consequence of this
is that assignment is very efficient.
Iterating and Copying Collections
When we assign large collections, such as long lists, the savings are very
apparent. Here is an example:
>>> songs = ["Because", "Boys", "Carol"]
>>> beatles = songs
>>> beatles, songs
(['Because', 'Boys', 'Carol'], ['Because', 'Boys', 'Carol'])
Here, a new object reference (beatles) has been created, and both object
references refer to the same list—no copying has taken place.
Since lists are mutable, we can apply a change. For example:
>>> beatles[2] = "Cayenne"
>>> beatles, songs
(['Because', 'Boys', 'Cayenne'], ['Because', 'Boys', 'Cayenne'])
We applied the change using the beatles variable—but this is an object reference referring to the same list as songs refers to. So any change made through
either object reference is visible to the other. This is most often the behavior
we want, since copying large collections is potentially expensive. It also means,
for example, that we can pass a list or other mutable collection data type as an
argument to a function, modify the collection in the function, and know that the
modified collection will be accessible after the function call has completed.
However, in some situations, we really do want a separate copy of the collection
(or other mutable object). For sequences, when we take a slice—for example,
songs[:2]—the slice is always an independent copy of the items copied. So to
copy an entire sequence we can do this:
>>> songs = ["Because", "Boys", "Carol"]
>>> beatles = songs[:]
>>> beatles[2] = "Cayenne"
>>> beatles, songs
(['Because', 'Boys', 'Cayenne'], ['Because', 'Boys', 'Carol'])
For dictionaries and sets, copying can be achieved using dict.copy() and
set.copy(). In addition, the copy module provides the copy.copy() function that
returns a copy of the object it is given. Another way to copy the built-in collection types is to use the type as a function with the collection to be copied as its
argument. Here are some examples:
copy_of_dict_d = dict(d)
copy_of_list_L = list(L)
copy_of_set_s = set(s)
Note, though, that all of these copying techniques are shallow—that is, only
object references are copied and not the objects themselves. For immutable
Chapter 3. Collection Data Types
data types like numbers and strings this has the same effect as copying (except
that it is more efficient), but for mutable data types such as nested collections
this means that the objects they refer to are referred to both by the original
collection and by the copied collection. The following snippet illustrates this:
>>> x = [53, 68, ["A", "B", "C"]]
>>> y = x[:] # shallow copy
>>> x, y
([53, 68, ['A', 'B', 'C']], [53, 68, ['A', 'B', 'C']])
>>> y[1] = 40
>>> x[2][0] = 'Q'
>>> x, y
([53, 68, ['Q', 'B', 'C']], [53, 40, ['Q', 'B', 'C']])
When list x is shallow-copied, the reference to the nested list ["A", "B", "C"] is
copied. This means that both x and y have as their third item an object reference that refers to this list, so any changes to the nested list are seen by both x
and y. If we really need independent copies of arbitrarily nested collections, we
can deep-copy:
>>> import copy
>>> x = [53, 68, ["A", "B", "C"]]
>>> y = copy.deepcopy(x)
>>> y[1] = 40
>>> x[2][0] = 'Q'
>>> x, y
([53, 68, ['Q', 'B', 'C']], [53, 40, ['A', 'B', 'C']])
Here, lists x and y, and the list items they contain, are completely independent.
Note that from now on we will use the terms copy and shallow copy
interchangeably—if we mean deep copy, we will say so explicitly.
We have now completed our review of Python’s built-in collection data types,
and three of the standard library collection types (collections.namedtuple,
collections.defaultdict, and collections.OrderedDict). Python also provides
the collections.deque type, a double-ended queue, and many other collection
types are available from third parties and from the Python Package Index, But now we will look at a couple of slightly longer examples that draw together many of the things covered in this chapter, and in the
preceding one.
The first program is about seventy lines long and involves text processing. The
second program is around ninety lines long and is mathematical in flavor. Between them, the programs make use of dictionaries, lists, named tuples, and
sets, and both make great use of the str.format() method from the preceding
Imagine we are setting up a new computer system and need to generate usernames for all of our organization’s staff. We have a plain text data file (UTF8 encoding) where each line represents a record and fields are colon-delimited.
Each record concerns one member of the staff and the fields are their unique
staff ID, forename, middle name (which may be an empty field), surname,
and department name. Here is an extract of a few lines from an example
data/users.txt data file:
The program must read in all the data files given on the command line, and for
every line (record) must extract the fields and return the data with a suitable
username. Each username must be unique and based on the person’s name.
The output must be text sent to the console, sorted alphabetically by surname
and forename, for example:
-------------------------------Landale, Nadelle................
Montgomery, Albert L............
Montgomery, Albert L............
Each record has exactly five fields, and although we could refer to them by
number, we prefer to use names to keep our code clear:
It is a Python convention that identifiers written in all uppercase characters
are to be treated as constants.
We also need to create a named tuple type for holding the data on each user:
User = collections.namedtuple("User",
"username forename middlename surname id")
We will see how the constants and the User named tuple are used when we look
at the rest of the code.
Chapter 3. Collection Data Types
The program’s overall logic is captured in the main() function:
def main():
if len(sys.argv) == 1 or sys.argv[1] in {"-h", "--help"}:
print("usage: {0} file1 [file2 [... fileN]]".format(
usernames = set()
users = {}
for filename in sys.argv[1:]:
for line in open(filename, encoding="utf8"):
line = line.rstrip()
if line:
user = process_line(line, usernames)
users[(user.surname.lower(), user.forename.lower(),] = user
If the user doesn’t provide any filenames on the command line, or if they type
“-h” or “--help” on the command line, we simply print a usage message and
terminate the program.
For each line read, we strip off any trailing whitespace (e.g., \n) and process
only nonempty lines. This means that if the data file contains blank lines they
will be safely ignored.
We keep track of all the allocated usernames in the usernames set to ensure that
we don’t create any duplicates. The data itself is held in the users dictionary,
with each user (member of the staff) stored as a dictionary item whose key is
a tuple of the user’s surname, forename, and ID, and whose value is a named
tuple of type User. Using a tuple of the user’s surname, forename, and ID for the
dictionary’s keys means that if we call sorted() on the dictionary, the iterable
returned will be in the order we want (i.e., surname, forename, ID), without us
having to provide a key function.
def process_line(line, usernames):
fields = line.split(":")
username = generate_username(fields, usernames)
user = User(username, fields[FORENAME], fields[MIDDLENAME],
fields[SURNAME], fields[ID])
return user
Since the data format for each record is so simple, and because we’ve already
stripped the trailing whitespace from the line, we can extract the fields simply
by splitting on the colons. We pass the fields and the usernames set to the
generate_username() function, and then we create an instance of the User named
tuple type which we then return to the caller (main()), which inserts the user
into the users dictionary, ready for printing.
If we had not created suitable constants to hold the index positions, we would
be reduced to using numeric indexes, for example:
user = User(username, fields[1], fields[2], fields[3], fields[0])
Although this is certainly shorter, it is poor practice. First it isn’t clear to
future maintainers what each field is, and second it is vulnerable to data file
format changes—if the order or number of fields in a record changes, this code
will break everywhere it is used. But by using named constants in the face of
changes to the record struture, we would have to change only the values of the
constants, and all uses of the constants would continue to work.
def generate_username(fields, usernames):
username = ((fields[FORENAME][0] + fields[MIDDLENAME][:1] +
fields[SURNAME]).replace("-", "").replace("'", ""))
username = original_name = username[:8].lower()
count = 1
while username in usernames:
username = "{0}{1}".format(original_name, count)
count += 1
return username
We make a first attempt at creating a username by concatenating the first letter of the forename, the first letter of the middle name, and the whole surname,
and deleting any hyphens or single quotes from the resultant string. The code
for getting the first letter of the middle name is quite subtle. If we had used
fields[MIDDLENAME][0] we would get an IndexError exception for empty middle
names. But by using a slice we get the first letter if there is one, or an empty
string otherwise.
Next we make the username lowercase and no more than eight characters long.
If the username is in use (i.e., it is in the usernames set), we try the username
with a “1” tacked on at the end, and if that is in use we try with a “2”, and so
on until we get one that isn’t in use. Then we add the username to the set of
usernames and return the username to the caller.
def print_users(users):
namewidth = 32
usernamewidth = 9
print("{0:<{nw}} {1:^6} {2:{uw}}".format(
"Name", "ID", "Username", nw=namewidth, uw=usernamewidth))
print("{0:-<{nw}} {0:-<6} {0:-<{uw}}".format(
"", nw=namewidth, uw=usernamewidth))
Chapter 3. Collection Data Types
for key in sorted(users):
user = users[key]
initial = ""
if user.middlename:
initial = " " + user.middlename[0]
name = "{0.surname}, {0.forename}{1}".format(user, initial)
print("{0:.<{nw}} ({}) {1.username:{uw}}".format(
name, user, nw=namewidth, uw=usernamewidth))
Once all the records have been processed, the print_users() function is called,
with the users dictionary passed as its parameter.
78 ➤
The first print() statement prints the column titles, and the second print()
statement prints hyphens under each title. This second statement’s str.
format() call is slightly subtle. The string we give to be printed is "", that is, the
empty string—we get the hyphens by printing the empty string padded with
hyphens to the given widths.
Next we use a for … in loop to print the details of each user, extracting the
key for each user’s dictionary item in sorted order. For convenience we create
the user variable so that we don’t have to keep writing users[key] throughout
the rest of the function. In the loop’s first call to str.format() we set the name
variable to the user’s name in surname, forename (and optional initial) form.
We access items in the user named tuple by name. Once we have the user’s
name as a single string we print the user’s details, constraining each column,
(name, ID, username) to the widths we want.
The complete program (which differs from what we have reviewed only
in that it has some initial comment lines and some imports) is in The program’s structure—read in a data file, process each
record, write output—is one that is very frequently used, and we will meet it
again in the next example.
Suppose we have a bunch of data files containing numbers relating to some
processing we have done, and we want to produce some basic statistics to
give us some kind of overview of the data. Each file uses plain text (ASCII
encoding) with one or more numbers per line (whitespace-separated).
Here is an example of the kind of output we want to produce:
std. dev.
= [5.00, 7.00, 50.00]
Here, we read 183 numbers, with 5, 7, and 50 occurring most frequently, and
with a sample standard deviation of 235.01.
The statistics themselves are held in a named tuple called Statistics:
Statistics = collections.namedtuple("Statistics",
"mean mode median std_dev")
The main() function also serves as an overview of the program’s structure:
def main():
if len(sys.argv) == 1 or sys.argv[1] in {"-h", "--help"}:
print("usage: {0} file1 [file2 [... fileN]]".format(
numbers = []
frequencies = collections.defaultdict(int)
for filename in sys.argv[1:]:
read_data(filename, numbers, frequencies)
if numbers:
statistics = calculate_statistics(numbers, frequencies)
print_results(len(numbers), statistics)
print("no numbers found")
We store all the numbers from all the files in the numbers list. To calculate the
mode (“most frequently occurring”) numbers, we need to know how many times
each number occurs, so we create a default dictionary using the int() factory
function, to keep track of the counts.
We iterate over each filename and read in its data. We pass the list and default
dictionary as additional parameters so that the read_data() function can
update them. Once we have read all the data, assuming some numbers were
successfully read, we call calculate_statistics(). This returns a named tuple
of type Statistics which we then use to print the results.
def read_data(filename, numbers, frequencies):
for lino, line in enumerate(open(filename, encoding="ascii"),
for x in line.split():
number = float(x)
frequencies[number] += 1
except ValueError as err:
print("{filename}:{lino}: skipping {x}: {err}".format(
Chapter 3. Collection Data Types
We split every line on whitespace, and for each item we attempt to convert it to
a float. If a conversion succeeds—as it will for integers and for floating-point
numbers in both decimal and exponential notations—we add the number to
the numbers list and update the frequencies default dictionary. (If we had used
a plain dict, the update code would have been frequencies[number] = frequencies.get(number, 0) + 1.)
Using str.
mapping unpacking
81 ➤
If a conversion fails, we output the line number (starting from line 1 as is traditional for text files), the text we attempted to convert, and the ValueError
exception’s error text. Rather than using positional arguments (e.g., .format(filename, lino, etc., or explicitly named arguments, .format(filename=filename, lino=lino, etc.), we have retrieved the names and values of the local
variables by calling locals() and used mapping unpacking to pass these as
key–value named arguments to the str.format() method.
def calculate_statistics(numbers, frequencies):
mean = sum(numbers) / len(numbers)
mode = calculate_mode(frequencies, 3)
median = calculate_median(numbers)
std_dev = calculate_std_dev(numbers, mean)
return Statistics(mean, mode, median, std_dev)
This function is used to gather all the statistics together. Because the mean
(“average”) is so easy to calculate, we do so directly here. For the other statistics
we call dedicated functions, and at the end we return a Statistics named tuple
object that contains the four statistics we have calculated.
def calculate_mode(frequencies, maximum_modes):
highest_frequency = max(frequencies.values())
mode = [number for number, frequency in frequencies.items()
if frequency == highest_frequency]
if not (1 <= len(mode) <= maximum_modes):
mode = None
return mode
There may be more than one most-frequently-occurring number, so in addition to the dictionary of frequencies, this function also requires the caller
to specify the maximum number of modes that are acceptable. (The calculate_statistics() function is the caller, and it specified a maximum of
three modes.)
The max() function is used to find the highest value in the frequencies dictionary. Then, we use a list comprehension to create a list of those modes whose
frequency equals the highest value. We can compare using operator == since
all the frequencies are integers.
If the number of modes is 0 or greater than the maximum modes that are
acceptable, a mode of None is returned; otherwise, a sorted list of the modes
is returned.
def calculate_median(numbers):
numbers = sorted(numbers)
middle = len(numbers) // 2
median = numbers[middle]
if len(numbers) % 2 == 0:
median = (median + numbers[middle - 1]) / 2
return median
The median (“middle value”) is the value that occurs in the middle if the
numbers are arranged in order—except when the number of numbers is even,
in which case the middle falls between two numbers, so in that case the median
is the mean of the two middle numbers.
We begin by sorting the numbers into ascending order. Then we use truncating
(integer) division to find the index position of the middle number, which we
extract and store as the median. If the number of numbers is even, we make
the median the mean of the two middle numbers.
def calculate_std_dev(numbers, mean):
total = 0
for number in numbers:
total += ((number - mean) ** 2)
variance = total / (len(numbers) - 1)
return math.sqrt(variance)
The sample standard deviation is a measure of dispersion, that is, how far the
numbers differ from the mean. This function calculates the sample standard
deviation using the formula s = ∑ (xn −−1x) , where x is each number, −x is the mean,
and n is the number of numbers.
def print_results(count, statistics):
real = "9.2f"
if statistics.mode is None:
modeline = ""
elif len(statistics.mode) == 1:
modeline = "mode
= {0:{fmt}}\n".format(
statistics.mode[0], fmt=real)
modeline = ("mode
= [" +
", ".join(["{0:.2f}".format(m)
for m in statistics.mode]) + "]\n")
Chapter 3. Collection Data Types
= {0:6}
= {mean:{fmt}}
= {median:{fmt}}
std. dev. = {std_dev:{fmt}}""".format(
count, modeline, fmt=real, **statistics._asdict()))
78 ➤
111 ➤
Using str.
mapping unpacking
81 ➤
Most of this function is concerned with formatting the modes list into the modeline string. If there are no modes, the mode line is not printed at all. If there
is one mode, the mode list has just one item (mode[0]) which is printed using
the same format as is used for the other statistics. If there are several modes,
we print them as a list with each one formatted appropriately. This is done by
using a list comprehension to produce a list of mode strings, and then joining
all the strings in the list together with “, ” in between each one. The printing
at the end is easy thanks to our use of a named tuple and its _asdict() method,
in conjunction with mapping unpacking. This lets us access the statistics in
the statistics object using names rather than numeric indexes, and thanks to
Python’s triple-quoted strings we can lay out the text to be printed in an understandable way. Recall that if we use mapping unpacking to pass arguments to
the str.format() method, it may be done only once and only at the end.
There is one subtle point to note. The modes are printed as format item {1},
which is followed by a backslash. The backslash escapes the newline, so if the
mode is the empty string no blank line will appear. And it is because we have
escaped the newline that we must put \n at the end of the modeline string if it
is not empty.
In this chapter we covered all of Python’s built-in collection types, and also a
couple of collection types from the standard library. We covered the collection
sequence types, tuple, collections.namedtuple, and list, which support the
same slicing and striding syntax as strings. The use of the sequence unpacking operator (*) was also covered, and brief mention was made of starred arguments in function calls. We also covered the set types, set and frozenset, and
the mapping types, dict and collections.defaultdict.
We saw how to use the named tuples provided by Python’s standard library to
create simple custom tuple data types whose items can be accessed by index
position, or more conveniently, by name. We also saw how to create “constants”
by using variables with all uppercase names.
In the coverage of lists we saw that everything that can be done to tuples can
be done to lists. And thanks to lists being mutable they offer considerably
more functionality than tuples. This includes methods that modify the list
(e.g., list.pop()), and the ability to have slices on the left-hand side of an assignment, to provide insertion, replacement, and deletion of slices. Lists are
ideal for holding sequences of items, especially if we need fast access by index
When we discussed the set and frozenset types, we noted that they may
contain only hashable items. Sets provide fast membership testing and are
useful for filtering out duplicate data.
Dictionaries are in some ways similar to sets—for example, their keys must
be hashable and are unique just like the items in a set. But dictionaries hold
key–value pairs, whose values can be of any type. The dictionary coverage
included the dict.get() and dict.setdefault() methods, and the coverage of
default dictionaries showed an alternative to using these methods. Like sets,
dictionaries provide very fast membership testing and fast access by key.
Lists, sets, and dictionaries all offer compact comprehension syntaxes that can
be used to create collections of these types from iterables (which themselves
can be comprehensions), and with conditions attached if required. The range()
and zip() functions are frequently used in the creation of collections, both in
conventional for … in loops and in comprehensions.
Items can be deleted from the mutable collection types using the relevant
methods, such as list.pop() and set.discard(), or using del, for example, del
d[k] to delete an item with key k from dictionary d.
Python’s use of object references makes assignment extremely efficient, but
it also means that objects are not copied when the assignment operator (=) is
used. We saw the differences between shallow and deep copying, and later on
saw how lists can be shallow-copied using a slice of the entire list, L[:], and how
dictionaries can be shallow-copied using the dict.copy() method. Any copyable
object can be copied using functions from the copy module, with copy.copy()
performing a shallow copy, and copy.deepcopy() performing a deep copy.
We introduced Python’s highly optimized sorted() function. This function is
used a lot in Python programming, since Python doesn’t provide any intrinsically ordered collection data types, so when we need to iterate over collections
in sorted order, we use sorted().
Python’s built-in collection data types—tuples, lists, sets, frozen sets, and
dictionaries—are sufficient in themselves for all purposes. Nonetheless, a few
additional collection types are available in the standard library, and many
more are available from third parties.
We often need to read in collections from files, or write collections to files. In
this chapter we focused just on reading and writing lines of text in our very
brief coverage of text file handling. Full coverage of file handling is given in
Chapter 7, and additional means of providing data persistence is covered in
Chapter 12.
Chapter 3. Collection Data Types
In the next chapter, we will look more closely at Python’s control structures,
and introduce one that we have not seen before. We will also look in more depth
at exception-handling and at some additional statements, such as assert, that
we have not yet covered. In addition, we will cover the creation of custom functions, and in particular we will look at Python’s incredibly versatile argumenthandling facilities.
1. Modify the program to use a default dictionary. This is
an easy change requiring an additional import, and changes to just two
other lines. A solution is provided in
2. Modify the program so that it outputs the words in frequency of occurrence order rather than in alphabetical order. You’ll need
to iterate over the dictionary’s items and create a tiny two-line function
to extract each item’s value and pass this function as sorted()’s key function. Also, the call to print() will need to be changed appropriately. This
isn’t difficult, but it is slightly subtle. A solution is provided in
3. Modify the program so that it prints the details of
two users per line, limiting names to 17 characters and outputting a form
feed character after every 64 lines, with the column titles printed at the
start of every page. Here’s a sample of the expected output:
----------------Aitkin, Shatha...
Allison, Karma...
Annie, Neervana..
----------------Alderson, Nicole.
Alwood, Kole E...
Apperson, Lucyann
This is challenging. You’ll need to keep the column titles in variables so
that they can be printed when needed, and you’ll need to tweak the format
specifications to accommodate the narrower names. One way to achieve
pagination is to write all the output items to a list and then iterate over
the list using striding to get the left- and right-hand items, and using zip()
to pair them up. A solution is provided in and
a longer sample data file is provided in data/users2.txt.
● Control Structures
● Exception Handling
● Custom Functions
Control Structures and
This chapter’s first two sections cover Python’s control structures, with the
first section dealing with branching and looping and the second section covering exception-handling. Most of the control structures and the basics of
exception-handling were introduced in Chapter 1, but here we give more complete coverage, including additional control structure syntaxes, and how to
raise exceptions and create custom exceptions.
The third and largest section is devoted to creating custom functions, with
detailed coverage of Python’s extremely versatile argument handling. Custom
functions allow us to package up and parameterize functionality—this reduces
the size of our code by eliminating code duplication and provides code reuse.
(In the following chapter we will see how to create custom modules so that we
can make use of our custom functions in multiple programs.)
Control Structures
Python provides conditional branching with if statements and looping with
while and for … in statements. Python also has a conditional expression—this
is a kind of if statement that is Python’s answer to the ternary operator (?:)
used in C-style languages.
Conditional Branching
As we saw in Chapter 1, this is the general syntax for Python’s conditional
branch statement:
if boolean_expression1:
Control Structures
The parentheses also make things clearer for human readers.
Conditional expressions can be used to improve messages printed for users.
For example, when reporting the number of files processed, instead of printing “0 file(s)”, “1 file(s)”, and similar, we could use a couple of conditional expressions:
print("{0} file{1}".format((count if count != 0 else "no"),
("s" if count != 1 else "")))
This will print “no files”, “1 file”, “2 files”, and similar, which gives a much more
professional impression.
Python provides a while loop and a for … in loop, both of which have a more
sophisticated syntax than the basics we showed in Chapter 1.
while Loops
Here is the complete general syntax of the while loop:
while boolean_expression:
The else clause is optional. As long as the boolean_expression is True, the while
block’s suite is executed. If the boolean_expression is or becomes False, the
loop terminates, and if the optional else clause is present, its suite is executed.
Inside the while block’s suite, if a continue statement is executed, control
is immediately returned to the top of the loop, and the boolean_expression is
evaluated again. If the loop does not terminate normally, any optional else
clause’s suite is skipped.
The optional else clause is rather confusingly named since the else clause’s
suite is always executed if the loop terminates normally. If the loop is broken
out of due to a break statement, or a return statement (if the loop is in a
function or method), or if an exception is raised, the else clause’s suite is not
executed. (If an exception occurs, Python skips the else clause and looks for
a suitable exception handler—this is covered in the next section.) On the plus
side, the behavior of the else clause is the same for while loops, for … in loops,
and try … except blocks.
Let’s look at an example of the else clause in action. The str.index() and
list.index() methods return the index position of a given string or item, or
raise a ValueError exception if the string or item is not found. The str.find()
Chapter 4. Control Structures and Functions
method does the same thing, but on failure, instead of raising an exception it
returns an index of -1. There is no equivalent method for lists, but if we wanted
a function that did this, we could create one using a while loop:
def list_find(lst, target):
index = 0
while index < len(lst):
if lst[index] == target:
index += 1
index = -1
return index
This function searches the given list looking for the target. If the target is
found, the break statement terminates the loop, causing the appropriate index
position to be returned. If the target is not found, the loop runs to completion
and terminates normally. After normal termination, the else suite is executed,
and the index position is set to -1 and returned.
for Loops
Like a while loop, the full syntax of the for … in loop also includes an optional
else clause:
for expression in iterable:
The expression is normally either a single variable or a sequence of variables,
usually in the form of a tuple. If a tuple or list is used for the expression, each
item is unpacked into the expression’s items.
If a continue statement is executed inside the for … in loop’s suite, control is
immediately passed to the top of the loop and the next iteration begins. If the
loop runs to completion it terminates, and any else suite is executed. If the
loop is broken out of due to a break statement, or a return statement (if the loop
is in a function or method), or if an exception is raised, the else clause’s suite
is not executed. (If an exception occurs, Python skips the else clause and looks
for a suitable exception handler—this is covered in the next section.)
139 ➤
Here is a for … in loop version of the list_find() function, and like the while
loop version, it shows the else clause in action:
def list_find(lst, target):
for index, x in enumerate(lst):
Control Structures
if x == target:
index = -1
return index
As this code snippet implies, the variables created in the for … in loop’s expression continue to exist after the loop has terminated. Like all local variables,
they cease to exist at the end of their enclosing scope.
Exception Handling
Python indicates errors and exceptional conditions by raising exceptions, although some third-party Python libraries use more old-fashioned techniques,
such as “error” return values.
Catching and Raising Exceptions
Exceptions are caught using try … except blocks, whose general syntax is:
except exception_group1 as variable1:
except exception_groupN as variableN:
There must be at least one except block, but both the else and the finally
blocks are optional. The else block’s suite is executed when the try block’s suite
has finished normally—but it is not executed if an exception occurs. If there
is a finally block, it is always executed at the end.
Each except clause’s exception group can be a single exception or a parenthesized tuple of exceptions. For each group, the as variable part is optional; if
used, the variable contains the exception that occurred, and can be accessed in
the exception block’s suite.
If an exception occurs in the try block’s suite, each except clause is tried in
turn. If the exception matches an exception group, the corresponding suite is
executed. To match an exception group, the exception must be of the same type
Chapter 4. Control Structures and Functions
Here is a final version of the list_find() function, this time using exceptionhandling:
def list_find(lst, target):
index = lst.index(target)
except ValueError:
index = -1
return index
Here, we have effectively used the try … except block to turn an exception
into a return value; the same approach can also be used to catch one kind of
exception and raise another instead—a technique we will see shortly.
Python also offers a simpler try … finally block which is sometimes useful:
No matter what happens in the try block’s suite (apart from the computer
or program crashing!), the finally block’s suite will be executed. The with
statement used with a context manager (both covered in Chapter 8) can be
used to achieve a similar effect to using a try … finally block.
One common pattern of use for try … except … finally blocks is for handling
file errors. For example, the program reads a list of filenames on
the command line, and for each one produces another file with the same name,
but with its extension changed to .nb, and with the same contents except for no
blank lines. Here’s the program’s read_data() function:
def read_data(filename):
lines = []
fh = None
fh = open(filename, encoding="utf8")
for line in fh:
if line.strip():
except (IOError, OSError) as err:
return []
if fh is not None:
return lines
Exception Handling
We set the file object, fh, to None because it is possible that the open() call will
fail, in which case nothing will be assigned to fh (so it will stay as None), and
an exception will be raised. If one of the exceptions we have specified occurs
(IOError or OSError), after printing the error message we return an empty list.
But note that before returning, the finally block’s suite will be executed, so the
file will be safely closed—if it had been successfully opened in the first place.
Notice also that if an encoding error occurs, even though we don’t catch the
relevant exception (UnicodeDecodeError), the file will still be safely closed. In
such cases the finally block’s suite is executed and then the exception is passed
up the call stack—there is no return value since the function finishes as a
result of the unhandled exception. And in this case, since there is no suitable
except block to catch encoding error exceptions, the program will terminate
and print a traceback.
Dealing with
➤ 415
We could have written the except clause slightly less verbosely:
except EnvironmentError as err:
return []
This works because EnvironmentError is the base class for both IOError and
In Chapter 8 we will show a slightly more compact idiom for ensuring that files
are safely closed, that does not require a finally block.
Raising Exceptions
➤ 369
Exceptions provide a useful means of changing the flow of control. We can
take advantage of this either by using the built-in exceptions, or by creating
our own, raising either kind when we want to. There are three syntaxes for
raising exceptions:
raise exception(args)
raise exception(args) from original_exception
When the first syntax is used the exception that is specified should be either
one of the built-in exceptions, or a custom exception that is derived from
Exception. If we give the exception some text as its argument, this text will be
output if the exception is printed when it is caught. The second syntax is a
variation of the first—the exception is raised as a chained exception (covered
in Chapter 9) that includes the original_exception exception, so this syntax
is used inside except suites. When the third syntax is used, that is, when no
exception is specified, raise will reraise the currently active exception—and if
there isn’t one it will raise a TypeError.
➤ 419
Chapter 4. Control Structures and Functions
Custom Exceptions
Custom exceptions are custom data types (classes). Creating classes is covered
in Chapter 6, but since it is easy to create simple custom exception types, we
will show the syntax here:
class exceptionName(baseException): pass
The base class should be Exception or a class that inherits from Exception.
One use of custom exceptions is to break out of deeply nested loops. For
example, if we have a table object that holds records (rows), which hold fields
(columns), which have multiple values (items), we could search for a particular
value with code like this:
found = False
for row, record in enumerate(table):
for column, field in enumerate(record):
for index, item in enumerate(field):
if item == target:
found = True
if found:
if found:
if found:
print("found at ({0}, {1}, {2})".format(row, column, index))
print("not found")
The 15 lines of code are complicated by the fact that we must break out of each
loop separately. An alternative solution is to use a custom exception:
class FoundException(Exception): pass
for row, record in enumerate(table):
for column, field in enumerate(record):
for index, item in enumerate(field):
if item == target:
raise FoundException()
except FoundException:
print("found at ({0}, {1}, {2})".format(row, column, index))
print("not found")
Exception Handling
This cuts the code down to ten lines, or 11 including defining the exception,
and is much easier to read. If the item is found we raise our custom exception
and the except block’s suite is executed—and the else block is skipped. And if
the item is not found, no exception is raised and so the else suite is executed at
the end.
Let’s look at another example to see some of the different ways that exceptionhandling can be done. All of the snippets are taken from the program, a program that reads all the HTML files it is given on the command line
and performs some simple tests to verify that tags begin with “<” and end with
“>”, and that entities are correctly formed. The program defines four custom
InvalidEntityError(Exception): pass
InvalidNumericEntityError(InvalidEntityError): pass
InvalidAlphaEntityError(InvalidEntityError): pass
InvalidTagContentError(Exception): pass
The second and third exceptions inherit from the first; we will see why this is
useful when we discuss the code that uses the exceptions. The parse() function
that uses the exceptions is more than 70 lines long, so we will show only those
parts that are relevant to exception-handling.
fh = None
fh = open(filename, encoding="utf8")
errors = False
for lino, line in enumerate(fh, start=1):
for column, c in enumerate(line, start=1):
The code begins conventionally enough, setting the file object to None and
putting all the file handling in a try block. The program reads the file line by
line and reads each line character by character.
Notice that we have two try blocks; the outer one is used to handle file object
exceptions, and the inner one is used to handle parsing exceptions.
elif state == PARSING_ENTITY:
if c == ";":
if entity.startswith("#"):
if frozenset(entity[1:]) - HEXDIGITS:
raise InvalidNumericEntityError()
elif not entity.isalpha():
raise InvalidAlphaEntityError()
Chapter 4. Control Structures and Functions
The function has various states, for example, after reading an ampersand
(&), it enters the PARSING_ENTITY state, and stores the characters between (but
excluding) the ampersand and semicolon in the entity string.
set type
121 ➤
The part of the code shown here handles the case when a semicolon has been
found while reading an entity. If the entity is numeric (of the form “&#”, with
hexadecimal digits, and then “;”, for example, “&#20AC;”), we convert the
numeric part of it into a set and take away from the set all the hexadecimal
digits; if anything is left at least one invalid character was present and we
raise a custom exception. If the entity is alphabetic (of the form “&”, with
letters, and then“;”, for example, “&copy;”), we raise a custom exception if any
of its letters is not alphabetic.
except (InvalidEntityError,
InvalidTagContentError) as err:
if isinstance(err, InvalidNumericEntityError):
error = "invalid numeric entity"
elif isinstance(err, InvalidAlphaEntityError):
error = "invalid alphabetic entity"
elif isinstance(err, InvalidTagContentError):
error = "invalid tag"
print("ERROR {0} in {1} on line {2} column {3}"
.format(error, filename, lino, column))
if skip_on_first_error:
If a parsing exception is raised we catch it in this except block. By using the
InvalidEntityError base class, we catch both InvalidNumericEntityError and
InvalidAlphaEntityError exceptions. We then use isinstance() to check which
type of exception occurred, and to set the error message accordingly. The
built-in isinstance() function returns True if its first argument is the same type
as the type (or one of that type’s base types) given as its second argument.
We could have used a separate except block for each of the three custom
parsing exceptions, but in this case combining them means that we avoided
repeating the last four lines (from the print() call to raise), in each one.
The program has two modes of use. If skip_on_first_error is False, the program continues checking a file even after a parsing error has occurred;
this can lead to multiple error messages being output for each file. If
skip_on_first_error is True, once a parsing error has occurred, after the (one
and only) error message is printed, raise is called to reraise the parsing exception and the outer (per-file) try block is left to catch it.
➤ 242
Exception Handling
elif state == PARSING_ENTITY:
raise EOFError("missing ';' at end of " + filename)
At the end of parsing a file, we need to check to see whether we have been left in
the middle of an entity. If we have, we raise an EOFError, the built-in end-of-file
exception, but give it our own message text. We could just as easily have raised
a custom exception.
except (InvalidEntityError, InvalidTagContentError):
pass # Already handled
except EOFError as err:
print("ERROR unexpected EOF:", err)
except EnvironmentError as err:
if fh is not None:
For the outer try block we have used separate except blocks since the behavior
we want varies. If we have a parsing exception, we know that an error message
has already been output and the purpose is simply to break out of reading the
file and to move on to the next file, so we don’t need to do anything in the exception handler. If we get an EOFError it could be caused by a genuine premature end of file or it could be the result of us raising the exception ourselves.
In either case, we print an error message, and the exception’s text. If an EnvironmentError occurs (i.e., if an IOError or an OSError occurs), we simply print its
message. And finally, no matter what, if the file was opened, we close it.
Custom Functions
Functions are a means by which we can package up and parameterize functionality. Four kinds of functions can be created in Python: global functions, local
functions, lambda functions, and methods.
Every function we have created so far has been a global function. Global
objects (including functions) are accessible to any code in the same module
(i.e., the same .py file) in which the object is created. Global objects can also be
accessed from other modules, as we will see in the next chapter.
Local functions (also called nested functions) are functions that are defined
inside other functions. These functions are visible only to the function where
they are defined; they are especially useful for creating small helper functions
that have no use elsewhere. We first show them in Chapter 7.
Chapter 4. Control Structures and Functions
Online Documentation
Although this book provides solid coverage of the Python 3 language and
the built-in functions and most commonly used modules in the standard
library, Python’s online documentation provides a considerable amount
of reference documentation, both on the language, and particularly on
Python’s extensive standard library. The documentation is available online
at and is also provided with Python itself.
On Windows the documentation is supplied in the Windows help file format.
Click Start→All Programs→Python 3.x →Python Manuals to launch the Windows
help browser. This tool has both an Index and a Search function that makes
finding documentation easy. Unix users have the documentation in HTML
format. In addition to the hyperlinks, there are various index pages. There
is also a very convenient Quick Search function available on the left-hand side
of each page.
The most frequently used online document for new users is the Library
Reference, and for experienced users the Global Module Index. Both of
these have links to pages covering Python’s entire standard library—and
in the case of the Library Reference, links to pages covering all of Python’s
built-in functionality as well.
It is well worth skimming through the documentation, particularly the Library Reference or the Global Module Index, to see what Python’s standard
library offers, and clicking through to the documentation of whichever topics are of interest. This should provide an initial impression of what is available and should also help you to establish a mental picture of where you can
find the documentation you are interested in. (A brief summary of Python’s
standard library is provided in Chapter 5.)
Help is also available from the interpreter itself. If you call the builtin help() function with no arguments, you will enter the online help
system—simply follow the instructions to get the information you want,
and type “q” or “quit” to return to the interpreter. If you know what module
or data type you want help on, you can call help() with the module or data
type as its argument. For example, help(str) provides information on the str
data type, including all of its methods, help(dict.update) provides information on the dict collection data type’s update() method, and help(os) displays
information about the os module (providing it has been imported).
Once familiar with Python, it is often sufficient to just be reminded about
what attributes (e.g., what methods) a data type provides. This information
is available using the dir() function—for example, dir(str) lists all the
string methods, and dir(os) lists all the os module’s constants and functions
(again, providing the module has been imported).
Custom Functions
Lambda functions are expressions, so they can be created at their point of use;
however, they are much more limited than normal functions.
Methods are functions that are associated with a particular data type and
can be used only in conjunction with the data type—they are introduced in
Chapter 6 when we cover object-oriented programming.
Python provides many built-in functions, and the standard library and thirdparty libraries add hundreds more (thousands if we count all the methods), so
in many cases the function we want has already been written. For this reason,
it is always worth checking Python’s online documentation to see what is already available. See the sidebar “Online Documentation” (172 ➤).
The general syntax for creating a (global or local) function is:
def functionName(parameters):
The parameters are optional, and if there is more than one they are written as a
sequence of comma-separated identifiers, or as a sequence of identifier=value
pairs as we will discuss shortly. For example, here is a function that calculates
the area of a triangle using Heron’s formula:
def heron(a, b, c):
s = (a + b + c) / 2
return math.sqrt(s * (s - a) * (s - b) * (s - c))
Inside the function, each parameter, a, b, and c, is initialized with the corresponding value that was passed as an argument. When the function is called,
we must supply all of the arguments, for example, heron(3, 4, 5). If we give too
few or too many arguments, a TypeError exception will be raised. When we do
a call like this we are said to be using positional arguments, because each argument passed is set as the value of the parameter in the corresponding position.
So in this case, a is set to 3, b to 4, and c to 5, when the function is called.
Every function in Python returns a value, although it is perfectly acceptable
(and common) to ignore the return value. The return value is either a single
value or a tuple of values, and the values returned can be collections, so there
are no practical limitations on what we can return. We can leave a function at
any point by using the return statement. If we use return with no arguments,
or if we don’t have a return statement at all, the function will return None.
(In Chapter 6 we will cover the yield statement which can be used instead of
return in certain kinds of functions.)
Some functions have parameters for which there can be a sensible default. For
example, here is a function that counts the letters in a string, defaulting to the
ASCII letters:
Chapter 4. Control Structures and Functions
def letter_count(text, letters=string.ascii_letters):
letters = frozenset(letters)
count = 0
for char in text:
if char in letters:
count += 1
return count
We have specified a default value for the letters parameter by using the
parameter=default syntax. This allows us to call letter_count() with just one
argument, for example, letter_count("Maggie and Hopey"). Here, inside the
function, letters will be the string that was given as the default value. But we
can still change the default, for example, using an extra positional argument,
letter_count("Maggie and Hopey", "aeiouAEIOU"), or using a keyword argument
(covered next), letter_count("Maggie and Hopey", letters="aeiouAEIOU").
The parameter syntax does not permit us to follow parameters with default
values with parameters that don’t have defaults, so def bad(a, b=1, c): won’t
work. On the other hand, we are not forced to pass our arguments in the
order they appear in the function’s definition—instead, we can use keyword
arguments, passing each argument in the form name=value.
Here is a tiny function that returns the string it is given, or if it is longer than
the specified length, it returns a shortened version with an indicator added:
def shorten(text, length=25, indicator="..."):
if len(text) > length:
text = text[:length - len(indicator)] + indicator
return text
Here are a few example calls:
shorten("The Silkie")
shorten(length=7, text="The Silkie")
shorten("The Silkie", indicator="&", length=7) #
shorten("The Silkie", 7, "&")
Because both length and indicator have default values, either or both can be
omitted entirely, in which case the default is used—this is what happens in
the first call. In the second call we use keyword arguments for both of the
specified parameters, so we can order them as we like. The third call mixes
both positional and keyword arguments. We used a positional first argument
(positional arguments must always precede keyword arguments), and then two
keyword arguments. The fourth call simply uses positional arguments.
The difference between a mandatory parameter and an optional parameter
is that a parameter with a default is optional (because Python can use the
default), and a parameter with no default is mandatory (because Python
Chapter 4. Control Structures and Functions
Using a conditional expression we can save a line of code for each parameter
that has a mutable default argument.
Names and Docstrings
Using good names for a function and its parameters goes a long way toward
making the purpose and use of the function clear to other programmers—and
to ourselves some time after we have created the function. Here are a few rules
of thumb that you might like to consider.
• Use a naming scheme, and use it consistently. In this book we use UPPERCASE for constants, TitleCase for classes (including exceptions), camelCase for GUI (Graphical User Interface) functions and methods (covered
in Chapter 15), and lowercase or lowercase_with_underscores for everything else.
• For all names, avoid abbreviations, unless they are both standardized and
widely used.
• Be proportional with variable and parameter names: x is a perfectly good
name for an x-coordinate and i is fine for a loop counter, but in general the
name should be long enough to be descriptive. The name should describe
the data’s meaning rather than its type (e.g., amount_due rather than money),
unless the use is generic to a particular type—see, for example, the text
parameter in the shorten() example (➤ 177).
• Functions and methods should have names that say what they do or
what they return (depending on their emphasis), but never how they do
it—since that might change.
Here are a few naming examples:
def find(l, s, i=0):
def linear_search(l, s, i=0):
def first_index_of(sorted_name_list, name, start=0):
All three functions return the index position of the first occurrence of a
name in a list of names, starting from the given starting index and using an
algorithm that assumes the list is already sorted.
The first one is bad because the name gives no clue as to what will be found,
and its parameters (presumably) indicate the required types (list, string, integer) without indicating what they mean. The second one is bad because the
function name describes the algorithm originally used—it might have been
changed since. This may not matter to users of the function, but it will probably confuse maintainers if the name implies a linear search, but the algorithm
implemented has been changed to a binary search. The third one is good
Custom Functions
cause the function name says what is returned, and the parameter names clearly indicate what is expected.
None of the functions have any way of indicating what happens if the name
isn’t found—do they return, say, -1, or do they raise an exception? Somehow
such information needs to be documented for users of the function.
We can add documentation to any function by using a docstring—this is simply
a string that comes immediately after the def line, and before the function’s
code proper begins. For example, here is the shorten() function we saw earlier,
but this time reproduced in full:
def shorten(text, length=25, indicator="..."):
"""Returns text or a truncated copy with the indicator added
text is any string; length is the maximum length of the returned
string (including any indicator); indicator is the string added at
the end to indicate that the text has been shortened
>>> shorten("Second Variety")
'Second Variety'
>>> shorten("Voices from the Street", 17)
'Voices from th...'
>>> shorten("Radio Free Albemuth", 10, "*")
'Radio Fre*'
if len(text) > length:
text = text[:length - len(indicator)] + indicator
return text
It is not unusual for a function or method’s documentation to be longer than the
function itself. One convention is to make the first line of the docstring a brief
one-line description, then have a blank line followed by a full description, and
then to reproduce some examples as they would appear if typed in interactively.
In Chapter 5 and Chapter 9 we will see how examples in function documentation can be used to provide unit tests.
Argument and Parameter Unpacking
114 ➤
We saw in the previous chapter that we can use the sequence unpacking operator (*) to supply positional arguments. For example, if we wanted to compute
the area of a triangle and had the lengths of the sides in a list, we could make
the call like this, heron(sides[0], sides[1], sides[2]), or simply unpack the list
and do the much simpler call, heron(*sides). And if the list (or other sequence)
has more items than the function has parameters, we can use slicing to extract
exactly the right number of arguments.
Chapter 4. Control Structures and Functions
We can also use the sequence unpacking operator in a function’s parameter
list. This is useful when we want to create functions that can take a variable
number of positional arguments. Here is a product() function that computes
the product of the arguments it is given:
def product(*args):
result = 1
for arg in args:
result *= arg
return result
This function has one parameter called args. Having the * in front means
that inside the function the args parameter will be a tuple with its items set to
however many positional arguments are given. Here are a few example calls:
product(1, 2, 3, 4)
product(5, 3, 8)
# args == (1, 2, 3, 4); returns: 24
# args == (5, 3, 8); returns: 120
# args == (11,); returns: 11
We can have keyword arguments following positional arguments, as this
function to calculate the sum of its arguments, each raised to the given power, shows:
def sum_of_powers(*args, power=1):
result = 0
for arg in args:
result += arg ** power
return result
The function can be called with just positional arguments, for example,
sum_of_powers(1, 3, 5), or with both positional and keyword arguments, for example, sum_of_powers(1, 3, 5, power=2).
It is also possible to use * as a “parameter” in its own right. This is used to
signify that there can be no positional arguments after the *, although keyword
arguments are allowed. Here is a modified version of the heron() function.
This time the function takes exactly three positional arguments, and has one
optional keyword argument.
def heron2(a, b, c, *, units="square meters"):
s = (a + b + c) / 2
area = math.sqrt(s * (s - a) * (s - b) * (s - c))
return "{0} {1}".format(area, units)
Here are a few example calls:
heron2(25, 24, 7)
# returns: '84.0 square meters'
heron2(41, 9, 40, units="sq. inches") # returns: '180.0 sq. inches'
Custom Functions
heron2(25, 24, 7, "sq. inches")
# WRONG! raises TypeError
In the third call we have attempted to pass a fourth positional argument, but
the * does not allow this and causes a TypeError to be raised.
By making the * the first parameter we can prevent any positional arguments
from being used, and force callers to use keyword arguments. Here is such a
(fictitious) function’s signature:
def print_setup(*, paper="Letter", copies=1, color=False):
We can call print_setup() with no arguments, and accept the defaults. Or we
can change some or all of the defaults, for example, print_setup(paper="A4",
color=True). But if we attempt to use positional arguments, for example,
print_setup("A4"), a TypeError will be raised.
Just as we can unpack a sequence to populate a function’s positional arguments, we can also unpack a mapping using the mapping unpacking operator,
asterisk asterisk (**).★ We can use ** to pass a dictionary to the print_setup()
function. For example:
options = dict(paper="A4", color=True)
Here the options dictionary’s key–value pairs are unpacked with each key’s
value being assigned to the parameter whose name is the same as the key. If
the dictionary contains a key for which there is no corresponding parameter,
a TypeError is raised. Any argument for which the dictionary has no corresponding item is set to its default value—but if there is no default, a TypeError
is raised.
We can also use the mapping unpacking operator with parameters. This allows
us to create functions that will accept as many keyword arguments as are given. Here is an add_person_details() function that takes Social Security number
and surname positional arguments, and any number of keyword arguments:
def add_person_details(ssn, surname, **kwargs):
print("SSN =", ssn)
surname =", surname)
for key in sorted(kwargs):
{0} = {1}".format(key, kwargs[key]))
This function could be called with just the two positional arguments, or with
additional information, for example, add_person_details(83272171, "Luther",
forename="Lexis", age=47). This provides us with a lot of flexibility. And we
As we saw in Chapter 2, when used as a binary operator, ** is the pow() operator.
Chapter 4. Control Structures and Functions
can of course accept both a variable number of positional arguments and a
variable number of keyword arguments:
def print_args(*args, **kwargs):
for i, arg in enumerate(args):
print("positional argument {0} = {1}".format(i, arg))
for key in kwargs:
print("keyword argument {0} = {1}".format(key, kwargs[key]))
This function just prints the arguments it is given. It can be called with no
arguments, or with any number of positional and keyword arguments.
Accessing Variables in the Global Scope
It is sometimes convenient to have a few global variables that are accessed by
various functions in the program. This is usually okay for “constants”, but is
not a good practice for variables, although for short one-off programs it isn’t
always unreasonable.
The program takes an optional language (“en” or “fr”) and a
number on the command line and outputs the names of each of the digits it is
given. So if it is invoked with “123” on the command line, it will output “one
two three”. The program has three global variables:
Language = "en"
ENGLISH = {0: "zero", 1: "one", 2: "two", 3: "three", 4: "four",
5: "five", 6: "six", 7: "seven", 8: "eight", 9: "nine"}
FRENCH = {0: "zéro", 1: "un", 2: "deux", 3: "trois", 4: "quatre",
5: "cinq", 6: "six", 7: "sept", 8: "huit", 9: "neuf"}
We have followed the convention that all uppercase variable names indicate
constants, and have set the default language to English. (Python does not
provide a direct way to create constants, instead relying on programmers to
respect the convention.) Elsewhere in the program we access the Language
variable, and use it to choose the appropriate dictionary to use:
def print_digits(digits):
dictionary = ENGLISH if Language == "en" else FRENCH
for digit in digits:
print(dictionary[int(digit)], end=" ")
When Python encounters the Language variable in this function it looks in the
local (function) scope and doesn’t find it. So it then looks in the global (.py file)
scope, and finds it there. The end keyword argument used with the first print()
call is explained in the sidebar “The print() Function” (➤ 181).
Chapter 4. Control Structures and Functions
“fr”, but the global Language variable used in the print_digits() function would
remain unchanged as “en”.
For nontrivial programs it is best not to use global variables except as constants, in which case there is no need to use the global statement.
Lambda Functions
Lambda functions are functions created using the following syntax:
lambda parameters: expression
The parameters are optional, and if supplied they are normally just commaseparated variable names, that is, positional arguments, although the complete
argument syntax supported by def statements can be used. The expression cannot contain branches or loops (although conditional expressions are allowed),
and cannot have a return (or yield) statement. The result of a lambda expression is an anonymous function. When a lambda function is called it returns the
result of computing the expression as its result. If the expression is a tuple it
should be enclosed in parentheses.
Here is a simple lambda function for adding an s (or not) depending on whether
its argument is 1:
s = lambda x: "" if x == 1 else "s"
The lambda expression returns an anonymous function which we assign to the
variable s. Any (callable) variable can be called using parentheses, so given the
count of files processed in some operation we could output a message using the
s() function like this: print("{0} file{1} processed".format(count, s(count))).
Lambda functions are often used as the key function for the built-in sorted()
function and for the list.sort() method. Suppose we have a list of elements
as 3-tuples of (group, number, name), and we wanted to sort this list in various
ways. Here is an example of such a list:
elements = [(2, 12, "Mg"), (1, 11, "Na"), (1, 3, "Li"), (2, 4, "Be")]
If we sort this list, we get this result:
[(1, 3, 'Li'), (1, 11, 'Na'), (2, 4, 'Be'), (2, 12, 'Mg')]
144 ➤
We saw earlier when we covered the sorted() function that we can provide a
key function to alter the sort order. For example, if we wanted to sort the list
by number and name, rather than the natural ordering of group, number, and
name, we could write a tiny function, def ignore0(e): return e[1], e[2], which
could be provided as the key function. Creating lots of little functions like this
can be inconvenient, so a frequently used alternative is a lambda function:
Generator functions
➤ 279
Custom Functions
elements.sort(key=lambda e: (e[1], e[2]))
Here the key function is lambda e: (e[1], e[2]) with e being each 3-tuple element in the list. The parentheses around the lambda expression are required
when the expression is a tuple and the lambda function is created as a function’s argument. We could use slicing to achieve the same effect:
elements.sort(key=lambda e: e[1:3])
A slightly more elaborate version gives us sorting in case-insensitive name,
number order:
elements.sort(key=lambda e: (e[2].lower(), e[1]))
Here are two equivalent ways to create a function that calculates the area of a
triangle using the conventional 21 × base × height formula:
area = lambda b, h: 0.5 * b * h
def area(b, h):
return 0.5 * b * h
We can call area(6, 5), whether we created the function using a lambda expression or using a def statement, and the result will be the same.
135 ➤
Another neat use of lambda functions is when we want to create default dictionaries. Recall from the previous chapter that if we access a default dictionary
using a nonexistent key, a suitable item is created with the given key and with
a default value. Here are a few examples:
minus_one_dict = collections.defaultdict(lambda: -1)
point_zero_dict = collections.defaultdict(lambda: (0, 0))
message_dict = collections.defaultdict(lambda: "No message available")
If we access the minus_one_dict with a nonexistent key, a new item will be created with the given key and with a value of -1. Similarly for the point_zero_dict
where the value will be the tuple (0, 0), and for the message_dict where the value will be the “No message available” string.
What happens if a function receives arguments with invalid data? What
happens if we make a mistake in the implementation of an algorithm and
perform an incorrect computation? The worst thing that can happen is that the
program executes without any (apparent) problem and no one is any the wiser.
One way to help avoid such insidious problems is to write tests—something we
will briefly look at in Chapter 5. Another way is to state the preconditions and
postconditions and to indicate an error if any of these are not met. Ideally, we
should use tests and also state preconditions and postconditions.
Chapter 4. Control Structures and Functions
Preconditions and postconditions can be specified using assert statements,
which have the syntax:
assert boolean_expression, optional_expression
If the boolean_expression evaluates to False an AssertionError exception is
raised. If the optional optional_expression is given, it is used as the argument
to the AssertionError exception—this is useful for providing error messages.
Note, though, that assertions are designed for developers, not end-users.
Problems that occur in normal program use such as missing files or invalid
command-line arguments should be handled by other means, such as providing
an error or log message.
Here are two new versions of the product() function. Both versions are equivalent in that they require that all the arguments passed to them are nonzero,
and consider a call with a 0 argument to be a coding error.
def product(*args): # pessimistic
assert all(args), "0 argument"
result = 1
for arg in args:
result *= arg
return result
def product(*args): # optimistic
result = 1
for arg in args:
result *= arg
assert result, "0 argument"
return result
The “pessimistic” version on the left checks all the arguments (or up to the first
0 argument) on every call. The “optimistic” version on the right just checks the
result; after all, if any argument was 0, then the result will be 0.
If one of these product() functions is called with a 0 argument an AssertionError exception will be raised, and output similar to the following will be written to the error stream (sys.stderr, usually the console):
Traceback (most recent call last):
File "", line 456, in <module>
x = product(1, 2, 0, 4, 8)
File "", line 452, in product
assert result, "0 argument"
AssertionError: 0 argument
Python automatically provides a traceback that gives the filename, function,
and line number, as well as the error message we specified.
Once a program is ready for public release (and of course passes all its tests and
does not violate any assertions), what do we do about the assert statements?
We can tell Python not to execute assert statements—in effect, to throw them
away at runtime. This can be done by running the program at the command
line with the -O option, for example, python -O Another approach
is to set the PYTHONOPTIMIZE environment variable to O. If the docstrings are of
Custom Functions
no use to our users (and normally they wouldn’t be), we can use the -OO option
which in effect strips out both assert statements and docstrings: Note that
there is no environment variable for setting this option. Some developers take
a simpler approach: They produce a copy of their program with all assert statements commented out, and providing this passes their tests, they release the
assertion-free version.
In this section we draw together some of the techniques covered in this chapter
and show them in the context of a complete example program.
Very small Web sites are often created and maintained by hand. One way
to make this slightly more convenient is to have a program that can generate skeleton HTML files that can later be fleshed out with content. The program is an interactive program that prompts the user
for various details and then creates a skeleton HTML file. The program’s main()
function has a loop so that users can create skeleton after skeleton, and it retains common data (e.g., copyright information) so that users don’t have to type
it in more than once. Here is a transcript of a typical interaction:
Make HTML Skeleton
your name (for copyright): Harold Pinter
copyright year [2008]: 2009
filename: career-synopsis
title: Career Synopsis
description (optional): synopsis of the career of Harold Pinter
a keyword (optional): playwright
a keyword (optional): actor
a keyword (optional): activist
a keyword (optional):
the stylesheet filename (optional): style
skeleton career-synopsis.html
Create another (y/n)? [y]:
Make HTML Skeleton
Enter your name (for copyright) [Harold Pinter]:
Enter copyright year [2009]:
Enter filename:
Create another (y/n)? [y]: n
Chapter 4. Control Structures and Functions
Notice that for the second skeleton the name and year had as their defaults
the values entered previously, so they did not need to be retyped. But no
default for the filename is provided, so when that was not given the skeleton
was cancelled.
Now that we have seen how the program is used, we are ready to study the
code. The program begins with two imports:
import datetime
import xml.sax.saxutils
The datetime module provides some simple functions for creating and datetime.time objects. The xml.sax.saxutils module has a useful
xml.sax.saxutils.escape() function that takes a string and returns an equivalent string with the special HTML characters (“&”, “<”, and “>”) in their escaped forms (“&amp;”, “&lt;”, and “&gt;”).
Three global strings are defined; these are used as templates.
COPYRIGHT_TEMPLATE = "Copyright (c) {0} {1}. All rights reserved."
STYLESHEET_TEMPLATE = ('<link rel="stylesheet" type="text/css" '
'media="all" href="{0}" />\n')
HTML_TEMPLATE = """<?xml version="1.0"?>
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN" \
<html xmlns="" lang="en" xml:lang="en">
<!-- {copyright} -->
<meta name="Description" content="{description}" />
<meta name="Keywords" content="{keywords}" />
<meta equiv="content-type" content="text/html; charset=utf-8" />
78 ➤
These strings will be used as templates in conjunction with the str.format()
method. In the case of HTML_TEMPLATE we have used names rather than index
positions for the field names, for example, {title}. We will see shortly that we
must use keyword arguments to provide values for these.
class CancelledError(Exception): pass
One custom exception is defined; we will see it in use when we look at a couple
of the program’s functions.
The program’s main() function is used to set up some initial information, and
to provide a loop. On each iteration the user has the chance to enter some
information for the HTML page they want generated, and after each one they
are given the chance to finish.
def main():
information = dict(name=None,,
filename=None, title=None, description=None,
keywords=None, stylesheet=None)
while True:
print("\nMake HTML Skeleton\n")
except CancelledError:
if (get_string("\nCreate another (y/n)?", default="y").lower()
not in {"y", "yes"}):
The function returns a object that holds today’s date. We want just the year attribute. All the other items of information
are set to None since there are no sensible defaults that can be set.
Inside the while loop the program prints a title, then calls the populate_information() function with the information dictionary. This dictionary is updated
inside the populate_information() function. Next, the make_html_skeleton()
function is called—this function takes a number of arguments, but rather than
give explicit values for each one we have simply unpacked the information dictionary.
If the user cancels, for example, by not providing mandatory information,
the program prints out “Cancelled”. At the end of each iteration (whether
cancelled or not), the user is asked whether they want to create another
skeleton—if they don’t, we break out of the loop and the program terminates.
def populate_information(information):
name = get_string("Enter your name (for copyright)", "name",
if not name:
raise CancelledError()
year = get_integer("Enter copyright year", "year",
information["year"], 2000, + 1, True)
title = xml.sax.saxutils.escape(title)
description = xml.sax.saxutils.escape(description)
keywords = ",".join([xml.sax.saxutils.escape(k)
for k in keywords]) if keywords else ""
stylesheet = (STYLESHEET_TEMPLATE.format(stylesheet)
if stylesheet else "")
html = HTML_TEMPLATE.format(**locals())
To get the copyright text we call str.format() on the COPYRIGHT_TEMPLATE, supplying the year and name (suitably HTML-escaped) as positional arguments
to replace {0} and {1}. For the title and description we produce HTML-escaped
copies of their texts.
78 ➤
For the HTML keywords we have two cases to deal with, and we distinguish
them using a conditional expression. If no keywords have been entered, we set
the keywords string to be the empty using. Otherwise, we use a list comprehension to iterate over all the keywords to produce a new list of strings, with each
one being HTML-escaped. This list is then joined into a single string with a
comma separating each item using str.join().
The stylesheet text is created in a similar way to the copyright text, but within
the context of a conditional expression so that the text is the empty string if
no stylesheet is specified.
Using str.
mapping unpacking
81 ➤
The html text is created from the HTML_TEMPLATE, with keyword arguments used
to provide the data for the replacement fields rather than the positional arguments used for the other template strings. Rather than pass each argument
explicitly using key=value syntax, we have used mapping unpacking on the
mapping returned by locals() to do this for us. (The alternative would be to
write the format() call as .format(title=title, copyright=copyright, etc.)
fh = None
fh = open(filename, "w", encoding="utf8")
except EnvironmentError as err:
print("ERROR", err)
print("Saved skeleton", filename)
if fh is not None:
Once the HTML has been prepared we write it to the file with the given
filename. We inform the user that the skeleton has been saved—or of the error
message if something went wrong. As usual we use a finally clause to ensure
that the file is closed if it was opened.
Chapter 4. Control Structures and Functions
def get_string(message, name="string", default=None,
minimum_length=0, maximum_length=80):
message += ": " if default is None else " [{0}]: ".format(default)
while True:
line = input(message)
if not line:
if default is not None:
return default
if minimum_length == 0:
return ""
raise ValueError("{0} may not be empty".format(
if not (minimum_length <= len(line) <= maximum_length):
raise ValueError("{name} must have at least "
"{minimum_length} and at most "
"{maximum_length} characters".format(
return line
except ValueError as err:
print("ERROR", err)
Using str.
mapping unpacking
81 ➤
This function has one mandatory argument, message, and four optional arguments. If a default value is given we include it in the message string so that
the user can see the default they would get if they just press Enter without typing any text. The rest of the function is enclosed in an infinite loop. The loop
can be broken out of by the user entering a valid string—or by accepting the
default (if given) by just pressing Enter. If the user makes a mistake, an error
message is printed and the loop continues. As usual, rather than explicitly using key=value syntax to pass local variables to str.format() with a format string
that uses named fields, we have simply used mapping unpacking on the mapping returned by locals() to do this for us.
The user could also break out of the loop, and indeed out of the entire program,
by typing Ctrl+C—this would cause a KeyboardInterrupt exception to be raised,
and since this is not handled by any of the program’s exception handlers, would
cause the program to terminate and print a traceback. Should we leave such
a “loophole”? If we don’t, and there is a bug in our program, we could leave the
user stuck in an infinite loop with no way out except to kill the process. Unless
there is a very strong reason to prevent Ctrl+C from terminating a program, it
should not be caught by any exception handler.
Notice that this function is not specific to the
program—it could be reused in many interactive programs of this type. Such
reuse could be achieved by copying and pasting, but that would lead to
tenance headaches—in the next chapter we will see how to create custom modules with functionality that can be shared across any number of programs.
def get_integer(message, name="integer", default=None, minimum=0,
maximum=100, allow_zero=True):
This function is so similar in structure to the get_string() function that it
would add nothing to reproduce it here. (It is included in the source code that
accompanies the book, of course.) The allow_zero parameter can be useful
when 0 is not a valid value but where we want to permit one invalid value to
signify that the user has cancelled. Another approach would be to pass an
invalid default value, and if that is returned, take it to mean that the user
has cancelled.
The last statement in the program is simply a call to main(). Overall the program is slightly more than 150 lines and shows several features of the Python
language introduced in this chapter and the previous ones.
This chapter covered the complete syntax for all of Python’s control structures.
It also showed how to raise and catch exceptions, and how to create custom
exception types.
Most of the chapter was devoted to custom functions. We saw how to create
functions and presented some rules of thumb for naming functions and their
parameters. We also saw how to provide documentation for functions. Python’s
versatile parameter syntax and argument passing were covered in detail, including both fixed and variable numbers of positional and keyword arguments,
and default values for arguments of both immutable and mutable data types.
We also briefly recapped sequence unpacking with * and showed how to do
mapping unpacking with **. Mapping unpacking is particularly useful when
applied to a dictionary (or other mapping), or to the mapping returned by locals(), for passing key–value arguments to a str.format() format string that
uses named fields.
If we need to assign a new value to a global variable inside a function, we can
do so by declaring that the variable is global, thereby preventing Python from
creating a local variable and assigning to that. In general, though, it is best to
use global variables only for constants.
Lambda functions are often used as key functions, or in other contexts where
functions must be passed as parameters. This chapter showed how to create
lambda functions, both as anonymous functions and as a means of creating
small named one-line functions by assigning them to a variable.
Chapter 4. Control Structures and Functions
The chapter also covered the use of the assert statement. This statement
is very useful for specifying the preconditions and postconditions that we
expect to be true on every use of a function, and can be a real aid to robust
programming and bug hunting.
In this chapter we covered all the fundamentals of creating functions, but
many other techniques are available to us. These include creating dynamic
functions (creating functions at runtime, possibly with implementations that
differ depending on circumstances), covered in Chapter 5; local (nested) functions, covered in Chapter 7; and recursive functions, generator functions, and
so on, covered in Chapter 8.
Although Python has a considerable amount of built-in functionality, and a
very extensive standard library, it is still likely that we will write some functions that would be useful in many of the programs we develop. Copying and
pasting such functions would lead to maintenance nightmares, but fortunately Python provides a clean easy-to-use solution: custom modules. In the next
chapter we will learn how to create our own modules with our own functions
inside them. We will also see how to import functionality from the standard
library and from our own modules, and will briefly review what the standard
library has to offer so that we can avoid reinventing the wheel.
Write an interactive program that maintains lists of strings in files.
When the program is run it should create a list of all the files in the current
directory that have the .lst extension. Use os.listdir(".") to get all the files
and filter out those that don’t have the .lst extension. If there are no matching
files the program should prompt the user to enter a filename—adding the .lst
extension if the user doesn’t enter it. If there are one or more .lst files they
should be printed as a numbered list starting from 1. The user should be asked
to enter the number of the file they want to load, or 0, in which case they should
be asked to give a filename for a new file.
If an existing file was specified its items should be read. If the file is empty, or
if a new file was specified, the program should show a message, “no items are
in the list”.
If there are no items, two options should be offered: “Add” and “Quit”. Once
the list has one or more items, the list should be shown with each item numbered from 1, and the options offered should be “Add”, “Delete”, “Save” (unless
already saved), and “Quit”. If the user chooses “Quit” and there are unsaved
changes they should be given the chance to save. Here is a transcript of a session with the program (with most blank lines removed, and without the “List
Keeper” title shown above the list each time):
Choose filename: movies
-- no items are in the list -[A]dd [Q]uit [a]: a
Add item: Love Actually
1: Love Actually
[A]dd [D]elete [S]ave
Add item: About a Boy
1: About a Boy
2: Love Actually
[A]dd [D]elete [S]ave
Add item: Alien
[Q]uit [a]: a
[Q]uit [a]:
1: About a Boy
2: Alien
3: Love Actually
[A]dd [D]elete [S]ave [Q]uit [a]: k
ERROR: invalid choice--enter one of 'AaDdSsQq'
Press Enter to continue...
[A]dd [D]elete [S]ave [Q]uit [a]: d
Delete item number (or 0 to cancel): 2
1: About a Boy
2: Love Actually
[A]dd [D]elete [S]ave [Q]uit [a]: s
Saved 2 items to movies.lst
Press Enter to continue...
1: About a Boy
2: Love Actually
[A]dd [D]elete [Q]uit [a]:
Add item: Four Weddings and a Funeral
1: About a Boy
2: Four Weddings and a Funeral
3: Love Actually
[A]dd [D]elete [S]ave [Q]uit [a]: q
Save unsaved changes (y/n) [y]:
Saved 3 items to movies.lst
Keep the main() function fairly small (less than 30 lines) and use it to provide
the program’s main loop. Write a function to get the new or existing filename
(and in the latter case to load the items), and a function to present the options and get the user’s choice of option. Also write functions to add an item,
delete an item, print a list (of either items or filenames), load the list, and
save the list. Either copy the get_string() and get_integer() functions from, or write your own versions.
Chapter 4. Control Structures and Functions
When printing the list or the filenames, print the item numbers using a field
width of 1 if there are less than ten items, of 2 if there are less than 100 items,
and of 3 otherwise.
Keep the items in case-insensitive alphabetical order, and keep track of
whether the list is “dirty” (has unsaved changes). Offer the “Save” option only
if the list is dirty and ask the user whether they want to save unsaved changes
when they quit only if the list is dirty. Adding or deleting an item will make
the list dirty; saving the list will make it clean again.
A model solution is provided in; it is less than 200 lines of code.
● Modules and Packages
● Overview of Python’s Standard
Whereas functions allow us to parcel up pieces of code so that they can be
reused throughout a program, modules provide a means of collecting sets of
functions (and as we will see in the next chapter, custom data types) together
so that they can be used by any number of programs. Python also has facilities
for creating packages—these are sets of modules that are grouped together,
usually because their modules provide related functionality or because they
depend on each other.
This chapter’s first section describes the syntaxes for importing functionality
from modules and packages—whether from the standard library, or from our
own custom modules and packages. The section then goes on to show how to
create custom packages and custom modules. Two custom module examples
are shown, the first introductory and the second illustrating how to handle
many of the practical issues that arise, such as platform independence and
172 ➤
The second section provides a brief overview of Python’s standard library. It is
important to be aware of what the library has to offer, since using predefined
functionality makes programming much faster than creating everything from
scratch. Also, many of the standard library’s modules are widely used, well
tested, and robust. In addition to the overview, a few small examples are used
to illustrate some common use cases. And cross-references are provided for
modules covered in other chapters.
Modules and Packages
A Python module, simply put, is a .py file. A module can contain any Python
code we like. All the programs we have written so far have been contained in a
single .py file, and so they are modules as well as programs. The key difference
Chapter 5. Modules
is that programs are designed to be run, whereas modules are designed to be
imported and used by programs.
Not all modules have associated .py files—for example, the sys module is built
into Python, and some modules are written in other languages (most commonly, C). However, much of Python’s library is written in Python, so, for example, if we write import collections we can create named tuples by calling
collections.namedtuple(), and the functionality we are accessing is in the module file. It makes no difference to our programs what language a module is written in, since all modules are imported and used in the
same way.
Several syntaxes can be used when importing. For example:
import importable
import importable1, importable2, ..., importableN
import importable as preferred_name
Here importable is usually a module such as collections, but could be a package
or a module in a package, in which case each part is separated with a dot (.),
for example, os.path. The first two syntaxes are the ones we use throughout
this book. They are the simplest and also the safest because they avoid the
possibility of having name conflicts, since they force us to always use fully
qualified names.
The third syntax allows us to give a name of our choice to the package or module we are importing—theoretically this could lead to name clashes, but in
practice the as syntax is used to avoid them. Renaming is particularly useful
when experimenting with different implementations of a module. For example, if we had two modules MyModuleA and MyModuleB that had the same API
(Application Programming Interface), we could write import MyModuleA as MyModule in a program, and later on seamlessly switch to using import MyModuleB as
Where should import statements go? It is common practice to put all the import
statements at the beginning of .py files, after the shebang line, and after the
module’s documentation. And as we said back in Chapter 1, we recommend
importing standard library modules first, then third-party library modules,
and finally our own modules.
Here are some other import syntaxes:
from importable import object as preferred_name
from importable import object1, object2, ..., objectN
from importable import (object1, object2, object3, object4, object5,
object6, ..., objectN)
from importable import *
➤ 199
Modules and Packages
These syntaxes can cause name conflicts since they make the imported objects
(variables, functions, data types, or modules) directly accessible. If we want
to use the from … import syntax to import lots of objects, we can use multiple
lines either by escaping each newline except the last, or by enclosing the object
names in parentheses, as the third syntax illustrates.
In the last syntax, the * means “import everything that is not private”, which in
practical terms means either that every object in the module is imported except
for those whose names begin with a leading underscore, or, if the module has
a global __all__ variable that holds a list of names, that all the objects named
in the __all__ variable are imported.
Here are a few import examples:
import os
# safe fully qualified access
import os.path as path
# risk of name collision with path
from os import path
# risk of name collision with path
from os.path import basename
# risk of name collision with basename
from os.path import *
# risk of many name collisions
The from importable import * syntax imports all the objects from the module (or
all the modules from the package)—this could be hundreds of names. In the
case of from os.path import *, almost 40 names are imported, including dirname,
exists, and split, any of which might be names we would prefer to use for our
own variables or functions.
For example, if we write from os.path import dirname, we can conveniently call
dirname() without qualification. But if further on in our code we write dirname
= ".", the object reference dirname will now be bound to the string "." instead of
to the dirname() function, so if we try calling dirname() we will get a TypeError
exception because dirname now refers to a string and strings are not callable.
In view of the potential for name collisions the import * syntax creates, some
programming teams specify in their guidelines that only the import importable
syntax may be used. However, certain large packages, particularly GUI
(Graphical User Interface) libraries, are often imported this way because they
have large numbers of functions and classes (custom data types) that can be
tedious to type out by hand.
A question that naturally arises is, how does Python know where to look for
the modules and packages that are imported? The built-in sys module has a
➤ 200
Modules and Packages
At every subsequent import of the module Python will detect that the module
has already been imported and will do nothing.
When Python needs a module’s byte-code compiled code, it generates it
automatically—this differs from, say, Java, where compiling to byte code must
be done explicitly. First Python looks for a file with the same name as the
module’s .py file but with the extension .pyo—this is an optimized byte-code
compiled version of the module. If there is no .pyo file (or if it is older than
the .py file, that is, if it is out of date), Python looks for a file with the extension .pyc—this is a nonoptimized byte-code compiled version of the module. If
Python finds an up-to-date byte-code compiled version of the module, it loads
it; otherwise, Python loads the .py file and compiles a byte-code compiled version. Either way, Python ends up with the module in memory in byte-code compiled form.
If Python had to byte-compile the .py file, it saves a .pyc version (or .pyo if -O
was specified on Python’s command line, or is set in the PYTHONOPTIMIZE environment variable), providing the directory is writable. Saving the byte code can
be avoided by using the -B command-line option, or by setting the PYTHONDONTWRITEBYTECODE environment variable.
Using byte-code compiled files leads to faster start-up times since the interpreter only has to load and run the code, rather than load, compile, (save if
possible), and run the code; runtimes are not affected, though. When Python is
installed, the standard library modules are usually byte-code compiled as part
of the installation process.
A package is simply a directory that contains a set of modules and a file called Suppose, for example, that we had a fictitious set of module files
for reading and writing various graphics file formats, such as,,,, and, all of which provided the functions load(), save(),
and so on.★ We could keep the modules in the same directory as our program,
but for a large program that uses scores of custom modules the graphics
modules will be dispersed. By putting them in their own subdirectory, say,
Graphics, they can be kept together. And if we put an empty file in
the Graphics directory along with them, the directory will become a package:
Extensive support for handling graphics files is provided by a variety of third-party modules,
most notably the Python Imaging Library (
Chapter 5. Modules
As long as the Graphics directory is a subdirectory inside our program’s directory or is in the Python path, we can import any of these modules and make use
of them. We must be careful to ensure that our top-level module name (Graphics) is not the same as any top-level name in the standard library so as to avoid
name conflicts. (On Unix this is easily done by starting with an uppercase letter since all of the standard library’s modules have lowercase names.) Here’s
how we can import and use our module:
import Graphics.Bmp
image = Graphics.Bmp.load("bashful.bmp")
For short programs some programmers prefer to use shorter names, and
Python makes this possible using two slightly different approaches.
import Graphics.Jpeg as Jpeg
image = Jpeg.load("doc.jpeg")
Here we have imported the Jpeg module from the Graphics package and told
Python that we want to refer to it simply as Jpeg rather than using its fully
qualified name, Graphics.Jpeg.
from Graphics import Png
image = Png.load("dopey.png")
This code snippet imports the Png module directly from the Graphics package.
This syntax (from … import) makes the Png module directly accessible.
We are not obliged to use the original package names in our code. For example:
from Graphics import Tiff as picture
image = picture.load("grumpy.tiff")
Here we are using the Tiff module, but have in effect renamed it inside our
program as the picture module.
In some situations it is convenient to load in all of a package’s modules using
a single statement. To do this we must edit the package’s file
to contain a statement which specifies which modules we want loaded. This
statement must assign a list of module names to the special variable __all__.
For example, here is the necessary line for the Graphics/ file:
__all__ = ["Bmp", "Jpeg", "Png", "Tiff", "Xpm"]
Modules and Packages
That is all that is required, although we are free to put any other code we like in
the file. Now we can write a different kind of import statement:
from Graphics import *
image = Xpm.load("sleepy.xpm")
The from package import * syntax directly imports all the modules named in the
__all__ list. So, after this import, not only is the Xpm module directly accessible,
but so are all the others.
As noted earlier, this syntax can also be applied to a module, that is, from module
import *, in which case all the functions, variables, and other objects defined in
the module (apart from those whose names begin with a leading underscore)
will be imported. If we want to control exactly what is imported when the from
module import * syntax is used, we can define an __all__ list in the module itself,
in which case doing from module import * will import only those objects named
in the __all__ list.
So far we have shown only one level of nesting, but Python allows us to nest
packages as deeply as we like. So we could have a subdirectory inside the
Graphics directory, say, Vector, with module files inside that, such as and
For the Vector directory to be a package it must have an file, and
as noted, this can be empty or could have an __all__ list as a convenience for
programmers who want to import using from Graphics.Vector import *.
To access a nested package we just build on the syntax we have already used:
import Graphics.Vector.Eps
image = Graphics.Vector.Eps.load("sneezy.eps")
The fully qualified name is rather long, so some programmers try to keep their
module hierarchies fairly flat to avoid this.
Chapter 5. Modules
import Graphics.Vector.Svg as Svg
image = Svg.load("snow.svg")
We can always use our own short name for a module, as we have done here,
although this does increase the risk of having a name conflict.
All the imports we have used so far (and that we will use throughout the rest
of the book) are absolute imports—this means that every module we import is
in one of sys.path’s directories (or subdirectories if the import name included
one or more periods which effectively serve as path separators). When creating
large multimodule multidirectory packages it is often useful to import other
modules that are part of the same package. For example, in or
we could get access to the Png module using a conventional import, or using a
relative import:
import Graphics.Png as Png
from ..Graphics import Png
These two code snippets are equivalent; they both make the Png module directly
available inside the module where they are used. But note that relative imports, that is, imports that use the from module import syntax with leading dots
in front of the module name (each dot representing stepping up one directory),
can be used only in modules that are inside a package. Using relative imports
makes it easier to rename the top-level package and prevents accidentally importing standard modules rather than our own inside packages.
Custom Modules
Since modules are just .py files they can be created without formality. In this
section we will look at two custom modules. The first module, TextUtil (in file, contains just three functions: is_balanced() which returns True
if the string it is passed has balanced parentheses of various kinds, shorten()
(shown earlier; 177 ➤), and simplify(), a function that can strip spurious
whitespace and other characters from a string. In the coverage of this module
we will also see how to execute the code in docstrings as unit tests.
The second module, CharGrid (in file, holds a grid of characters and
allows us to “draw” lines, rectangles, and text onto the grid and to render the
grid on the console. This module shows some techniques that we have not seen
before and is more typical of larger, more complex modules.
The TextUtil Module
The structure of this module (and most others) differs little from that of a
program. The first line is the shebang line, and then we have some comments
(typically the copyright and license information). Next it is common to have a
Modules and Packages
triple quoted string that provides an overview of the module’s contents, often
including some usage examples—this is the module’s docstring. Here is the
start of the file (but with the license comment lines omitted):
#!/usr/bin/env python3
# Copyright (c) 2008-9 Qtrac Ltd. All rights reserved.
This module provides a few string manipulation functions.
>>> is_balanced("(Python (is (not (lisp))))")
>>> shorten("The Crossing", 10)
'The Cro...'
>>> simplify(" some
with spurious
'some text with spurious whitespace'
import string
This module’s docstring is available to programs (or other modules) that import
the module as TextUtil.__doc__. After the module docstring come the imports,
in this case just one, and then the rest of the module.
177 ➤
We have already seen the shorten() function reproduced in full, so we will not
repeat it here. And since our focus is on modules rather than on functions,
although we will show the simplify() function in full, including its docstring,
we will show only the code for is_balanced().
This is the simplify() function, broken into two parts:
def simplify(text, whitespace=string.whitespace, delete=""):
r"""Returns the text with multiple spaces reduced to single spaces
The whitespace parameter is a string of characters, each of which
is considered to be a space.
If delete is not empty it should be a string, in which case any
characters in the delete string are excluded from the resultant
>>> simplify(" this
and\n that\t too")
'this and that too'
>>> simplify(" Washington
'Washington D.C.'
>>> simplify(" Washington
D.C.\n", delete=",;:.")
'Washington DC'
>>> simplify(" disemvoweled ", delete="aeiou")
67 ➤
Chapter 5. Modules
After the def line comes the function’s docstring, laid out conventionally with
a single line description, a blank line, further description, and then some
examples written as though they were typed in interactively. Because the
quoted strings are inside a docstring we must either escape the backslashes
inside them, or do what we have done here and use a raw triple quoted string.
result = []
word = ""
for char in text:
if char in delete:
elif char in whitespace:
if word:
word = ""
word += char
if word:
return " ".join(result)
The result list is used to hold “words”—strings that have no whitespace or
deleted characters. The given text is iterated over character by character, with
deleted characters skipped. If a whitespace character is encountered and a
word is in the making, the word is added to the result list and set to be an empty
string; otherwise, the whitespace is skipped. Any other character is added to
the word being built up. At the end a single string is returned consisting of all
the words in the result list joined with a single space between each one.
The is_balanced() function follows the same pattern of having a def line, then
a docstring with a single-line description, a blank line, further description,
and some examples, and then the code itself. Here is the code without the
def is_balanced(text, brackets="()[]{}<>"):
counts = {}
left_for_right = {}
for left, right in zip(brackets[::2], brackets[1::2]):
assert left != right, "the bracket characters must differ"
counts[left] = 0
left_for_right[right] = left
for c in text:
if c in counts:
counts[c] += 1
elif c in left_for_right:
left = left_for_right[c]
Modules and Packages
if counts[left] == 0:
return False
counts[left] -= 1
return not any(counts.values())
The function builds two dictionaries. The counts dictionary’s keys are the
opening characters (“(”, “[”, “{”, and “<”), and its values are integers. The
left_for_right dictionary’s keys are the closing characters (“)”, “]”, “}”, and “>”),
and its values are the corresponding opening characters. Once the dictionaries
are set up the function iterates character by character over the text. Whenever
an opening character is encountered, its corresponding count is incremented.
Similarly, when a closing character is encountered, the function finds out what
the corresponding opening character is. If the count for that character is 0 it
means we have reached one closing character too many so can immediately
return False; otherwise, the relevant count is decremented. At the end every
count should be 0 if all the pairs are balanced, so if any one of them is not 0 the
function returns False; otherwise, it returns True.
Up to this point everything has been much like any other .py file. If
was a program there would presumably be some more functions, and at the end
we would have a single call to one of those functions to start off the processing.
But since this is a module that is intended to be imported, defining functions is
sufficient. And now, any program or module can import TextUtil and make use
of it:
import TextUtil
text = " a
puzzling conundrum "
text = TextUtil.simplify(text) # text == 'a puzzling conundrum'
If we want the TextUtil module to be available to a particular program, we
just need to put in the same directory as the program. If we want to be available to all our programs, there are a few approaches that
can be taken. One approach is to put the module in the Python distribution’s
site-packages subdirectory—this is usually C:\Python31\Lib\site-packages on
Windows, but it varies on Mac OS X and other Unixes. This directory is in
the Python path, so any module that is here will always be found. A second
approach is to create a directory specifically for the custom modules we want
to use for all our programs, and to set the PYTHONPATH environment variable to
this directory. A third approach is to put the module in the local site-packages
subdirectory—this is %APPDATA%\Python\Python31\site-packages on Windows
and ~/.local/lib/python3.1/site-packages on Unix (including Mac OS X) and
is in the Python path. The second and third approaches have the advantage of
keeping our own code separate from the official installation.
Having the TextUtil module is all very well, but if we end up with lots of programs using it we might want to be more confident that it works as advertised.
Chapter 5. Modules
One really simple way to do this is to execute the examples in the docstrings
and make sure that they produce the expected results. This can be done by
adding just three lines at the end of the module’s .py file:
if __name__ == "__main__":
import doctest
Whenever a module is imported Python creates a variable for the module
called __name__ and stores the module’s name in this variable. A module’s
name is simply the name of its .py file but without the extension. So in this
example, when the module is imported __name__ will have the value "TextUtil",
and the if condition will not be met, so the last two lines will not be executed.
This means that these last three lines have virtually no cost when the module
is imported.
Whenever a .py file is run Python creates a variable for the program called
__name__ and sets it to the string "__main__". So if we were to run
as though it were a program, Python will set __name__ to "__main__" and the if
condition will evaluate to True and the last two lines will be executed.
The doctest.testmod() function uses Python’s introspection features to discover
all the functions in the module and their docstrings, and attempts to execute
all the docstring code snippets it finds. Running a module like this produces
output only if there are errors. This can be disconcerting at first since it doesn’t
look like anything happened at all, but if we pass a command-line flag of -v,
we will get output like this:
is_balanced("(Python (is (not (lisp))))")
simplify(" disemvoweled ", delete="aeiou")
4 items passed all tests:
3 tests in __main__
5 tests in __main__.is_balanced
3 tests in __main__.shorten
4 tests in __main__.simplify
15 tests in 4 items.
15 passed and 0 failed.
Test passed.
Modules and Packages
We have used an ellipsis to indicate a lot of lines that have been omitted. If
there are functions (or classes or methods) that don’t have tests, these are listed
when the -v option is used. Notice that the doctest module found the tests in
the module’s docstring as well as those in the functions’ docstrings.
Examples in docstrings that can be executed as tests are called doctests. Note
that when we write doctests, we are able to call simplify() and the other functions unqualified (since the doctests occur inside the module itself). Outside
the module, assuming we have done import TextUtil, we must use the qualified
names, for example, TextUtil.is_balanced().
In the next subsection we will see how to do more thorough tests—in particular,
testing cases where we expect failures, for example, invalid data causing exceptions. (Testing is covered more fully in Chapter 9.) We will also address some
other issues that arise when creating modules, including module initialization,
accounting for platform differences, and ensuring that if the from module import
* syntax is used, only the objects we want to be made public are actually imported into the importing program or module.
The CharGrid Module
The CharGrid module holds a grid of characters in memory. It provides functions for “drawing” lines, rectangles, and text on the grid, and for rendering the
grid onto the console. Here are the module’s docstring’s doctests:
resize(14, 50)
add_rectangle(0, 0, *get_size())
add_vertical_line(5, 10, 13)
add_vertical_line(2, 9, 12, "!")
add_horizontal_line(3, 10, 20, "+")
add_rectangle(0, 0, 5, 5, "%")
add_rectangle(5, 7, 12, 40, "#", True)
add_rectangle(7, 9, 10, 38, " ")
add_text(8, 10, "This is the CharGrid module")
add_text(1, 32, "Pleasantville", "@")
add_rectangle(6, 42, 11, 46, fill=True)
The CharGrid.add_rectangle() function takes at least four arguments, the topleft corner’s row and column and the bottom-right corner’s row and column.
The character used to draw the outline can be given as a fifth argument, and a
Boolean indicating whether the rectangle should be filled (with the same character as the outline) as a sixth argument. The first time we call it we pass the
third and fourth arguments by unpacking the 2-tuple (width, height), returned
by the CharGrid.get_size() function.
Chapter 5. Modules
By default, the CharGrid.render() function clears the screen before printing the
grid, but this can be prevented by passing False as we have done here. Here is
the grid that results from the preceding doctests:
@@@@@@@@@@@@@@@ *
@[email protected] *
@@@@@@@@@@@@@@@ *
################################# ****
## ****
## This is the CharGrid module ## ****
* !
## ****
* ! | ################################# ****
* ! | #################################
The module begins in the same way as the TextUtil module, with a shebang
line, copyright and license comments, and a module docstring that describes
the module and has the doctests quoted earlier. Then the code proper begins
with two imports, one of the sys module and the other of the subprocess module.
The subprocess module is covered more fully in Chapter 10.
The module has two error-handling policies in place. Several functions have
a char parameter whose actual argument must always be a string containing
exactly one character; a violation of this requirement is considered to be a fatal
coding error, so assert statements are used to verify the length. But passing
out-of-range row or column numbers is considered erroneous but normal, so
custom exceptions are raised when this happens.
We will now review some illustrative and key parts of the module’s code,
beginning with the custom exceptions:
class RangeError(Exception): pass
class RowRangeError(RangeError): pass
class ColumnRangeError(RangeError): pass
None of the functions in the module that raise an exception ever raise a
RangeError; they always raise the specific exception depending on whether an
out-of-range row or column was given. But by using a hierarchy, we give users
of the module the choice of catching the specific exception, or to catch either of
them by catching their RangeError base class. Note also that inside doctests the
exception names are used as they appear here, but if the module is imported
with import CharGrid, the exception names are, of course, CharGrid.RangeError,
CharGrid.RowRangeError, and CharGrid.ColumnRangeError.
Modules and Packages
_CHAR_ASSERT_TEMPLATE = ("char must be a single character: '{0}' "
"is too long")
_max_rows = 25
_max_columns = 80
_grid = []
_background_char = " "
Here we define some private data for internal use by the module. We use
leading underscores so that if the module is imported using from CharGrid
import *, none of these variables will be imported. (An alternative approach
would be to set an __all__ list.) The _CHAR_ASSERT_TEMPLATE is a string for use
with the str.format() function; we will see it used to give an error message in
assert statements. We will discuss the other variables as we encounter them.
if sys.platform.startswith("win"):
def clear_screen():["cmd.exe", "/C", "cls"])
def clear_screen():["clear"])
clear_screen.__doc__ = """Clears the screen using the underlying \
window system's clear screen command"""
The means of clearing the console screen is platform-dependent. On Windows
we must execute the cmd.exe program with appropriate arguments and on
most Unix systems we execute the clear program. The subprocess module’s function lets us run an external program, so we can use it
to clear the screen in the appropriate platform-specific way. The sys.platform
string holds the name of the operating system the program is running on, for
example, “win32” or “linux2”. So one way of handling the platform differences
would be to have a single clear_screen() function like this:
def clear_screen():
command = (["clear"] if not sys.platform.startswith("win") else
["cmd.exe", "/C", "cls"])
The disadvantage of this approach is that even though we know the platform
cannot change while the program is running, we perform the check every time
the function is called.
To avoid checking which platform the program is being run on every time
the clear_screen() function is called, we have created a platform-specific
clear_screen() function once when the module is imported, and from then on
we always use it. This is possible because the def statement is a Python statement like any other; when the interpreter reaches the if it executes either
the first or the second def statement, dynamically creating one or the other
Chapter 5. Modules
clear_screen() function. Since the function is not defined inside another func-
tion (or inside a class as we will see in the next chapter), it is still a global function, accessible like any other function in the module.
After creating the function we explicitly set its docstring; this avoids us having
to write the same docstring in two places, and also illustrates that a docstring
is simply one of the attributes of a function. Other attributes include the
function’s module and its name.
def resize(max_rows, max_columns, char=None):
"""Changes the size of the grid, wiping out the contents and
changing the background if the background char is not None
assert max_rows > 0 and max_columns > 0, "too small"
global _grid, _max_rows, _max_columns, _background_char
if char is not None:
assert len(char) == 1, _CHAR_ASSERT_TEMPLATE.format(char)
_background_char = char
_max_rows = max_rows
_max_columns = max_columns
_grid = [[_background_char for column in range(_max_columns)]
for row in range(_max_rows)]
This function uses an assert statement to enforce the policy that it is a coding
error to attempt to resize the grid smaller than 1 × 1. If a background character
is specified an assert is used to guarantee that it is a string of exactly one
character; if it is not, the assertion error message is the _CHAR_ASSERT_TEMPLATE’s
text with the {0} replaced with the given char string.
Unfortunately, we must use the global statement because we need to update a
number of global variables inside this function. This is something that using
an object-oriented approach can help us to avoid, as we will see in Chapter 6.
118 ➤
The _grid is created using a list comprehension inside a list comprehension.
Using list replication such as [[char] * columns] * rows will not work because
the inner list will be shared (shallow-copied). We could have used nested for …
in loops instead:
_grid = []
for row in range(_max_rows):
for column in range(_max_columns):
This code is arguably trickier to understand than the list comprehension, and
is much longer.
Modules and Packages
We will review just one of the drawing functions to give a flavor of how the
drawing is done, since our primary concern is with the implementation of the
module. Here is the add_horizontal_line() function, split into two parts:
def add_horizontal_line(row, column0, column1, char="-"):
"""Adds a horizontal line to the grid using the given char
>>> add_horizontal_line(8, 20, 25, "=")
>>> char_at(8, 20) == char_at(8, 24) == "="
>>> add_horizontal_line(31, 11, 12)
Traceback (most recent call last):
The docstring has two tests, one that is expected to work and another that is
expected to raise an exception. When dealing with exceptions in doctests the
pattern is to specify the “Traceback” line, since that is always the same and
tells the doctest module an exception is expected, then to use an ellipsis to
stand for the intervening lines (which vary), and ending with the exception line
we expect to get. The char_at() function is one of those provided by the module;
it returns the character at the given row and column position in the grid.
assert len(char) == 1, _CHAR_ASSERT_TEMPLATE.format(char)
for column in range(column0, column1):
_grid[row][column] = char
except IndexError:
if not 0 <= row <= _max_rows:
raise RowRangeError()
raise ColumnRangeError()
The code begins with the same character length check that is used in the resize() function. Rather than explicitly checking the row and column arguments, the function works by assuming that the arguments are valid. If an
IndexError exception occurs because a nonexistent row or column is accessed,
we catch the exception and raise the appropriate module-specific exception in
its place. This style of programming is known colloquially as “it’s easier to ask
forgiveness than permission”, and is generally considered more Pythonic (good
Python programming style) than “look before you leap”, where checks are made
in advance. Relying on exceptions to be raised rather than checking in advance
is more efficient when exceptions are rare. (Assertions don’t count as “look
before you leap” because they should never occur—and are often commented
out—in deployed code.)
Chapter 5. Modules
Almost at the end of the module, after all the functions have been defined,
there is a single call to resize():
resize(_max_rows, _max_columns)
This call initializes the grid to the default size (25 × 80) and ensures that code
that imports the module can safely make use of it immediately. Without this
call, every time the module was imported, the importing program or module
would have to call resize() to initialize the grid, forcing programmers to
remember that fact and also leading to multiple initializations.
if __name__ == "__main__":
import doctest
The last three lines of the module are the standard ones for modules that use
the doctest module to check their doctests. (Testing is covered more fully in
Chapter 9.)
The CharGrid module has an important failing: It supports only a single character grid. One solution to this would be to hold a collection of grids in the module, but that would mean that users of the module would have to provide a key
or index with every function call to identify which grid they were referring to.
In cases where multiple instances of an object are required, a better solution is
to create a module that defines a class (a custom data type), since we can create as many class instances (objects of the data type) as we like. An additional
benefit of creating a class is that we should be able to avoid using the global
statement by storing class (static) data. We will see how to create classes in the
next chapter.
Overview of Python’s Standard Library
Python’s standard library is generally described as “batteries included”, and
certainly a wide range of functionality is available, spread over around two
hundred packages and modules.
In fact, so many high-quality modules have been developed for Python over the
years, that to include them all in the standard library would probably increase
the size of the Python distribution packages by at least an order of magnitude.
So those modules that are in the library are more a reflection of Python’s history and of the interests of its core developers than of any concerted or systematic effort to create a “balanced” library. Also, some modules have proved
very difficult to maintain within the library—most notably the Berkeley DB
module—and so have been taken out of the library and are now maintained
independently. This means many excellent third-party modules are available
for Python that—despite their quality and usefulness—are not in the standard
Overview of Python’s Standard Library
library. (We will look at two such modules later on: the PyParsing and PLY
modules that are used to create parsers in Chapter 14.)
In this section we present a broad overview of what is on offer, taking a
thematic approach, but excluding those packages and modules that are of very
specialized interest and those which are platform-specific. In many cases a
small example is shown to give a flavor of some of the packages and modules;
cross-references are provided for those packages and modules that are covered
elsewhere in the book.
String Handling
The string module provides some useful constants such as string.ascii_letters and string.hexdigits. It also provides the string.Formatter class which we
can subclass to provide custom string formatters.★ The textwrap module can be
used to wrap lines of text to a specified width, and to minimize indentation.
The struct module provides functions for packing and unpacking numbers,
Booleans, and strings to and from bytes objects using their binary representations. This can be useful when handling data to be sent to or received from lowlevel libraries written in C. The struct and textwrap modules are used by the program covered in Chapter 7.
The difflib module provides classes and methods for comparing sequences,
such as strings, and is able to produce output both in standard “diff” formats
and in HTML.
Python’s most powerful string handling module is the re (regular expression)
module. This is covered in Chapter 13.
The io.StringIO class can provide a string-like object that behaves like an
in-memory text file. This can be convenient if we want to use the same code
that writes to a file to write to a string.
Example: The io.StringIO Class
Python provides two different ways of writing text to files. One way is to use
a file object’s write() method, and the other is to use the print() function
with the file keyword argument set to a file object that is open for writing.
For example:
print("An error message", file=sys.stdout)
sys.stdout.write("Another error message\n")
The term subclassing (or specializing) is used for when we create a custom data type (a class)
based on another class. Chapter 6 gives full coverage of this topic.
➤ 293
➤ 297
Chapter 5. Modules
Both lines of text are printed to sys.stdout, a file object that represents the
“standard output stream”—this is normally the console and differs from
sys.stderr, the “error output stream” only in that the latter is unbuffered.
(Python automatically creates and opens sys.stdin, sys.stdout, and sys.stderr
at program start-up.) The print() function adds a newline by default, although
we can stop this by giving the end keyword argument set to an empty string.
In some situations it is useful to be able to capture into a string the output
that is intended to go to a file. This can be achieved using the io.StringIO class
which provides an object that can be used just like a file object, but which holds
any data written to it in a string. If the io.StringIO object is given an initial
string, it can also be read as though it were a file.
We can access io.StringIO if we do import io, and we can use it to capture output
destined for a file object such as sys.stdout:
sys.stdout = io.StringIO()
If this line is put at the beginning of a program, after the imports but before
any use is made of sys.stdout, any text that is sent to sys.stdout will actually
be sent to the io.StringIO file-like object which this line has created and which
has replaced the standard sys.stdout file object. Now, when the print() and
sys.stdout.write() lines shown earlier are executed, their output will go to
the io.StringIO object instead of the console. (At any time we can restore the
original sys.stdout with the statement sys.stdout = sys.__stdout__.)
We can obtain all the strings that have been written to the io.StringIO object by calling the io.StringIO.getvalue() function, in this case by calling
sys.stdout.getvalue()—the return value is a string containing all the lines that
have been written. This string could be printed, or saved to a log or sent over
a network connection like any other string. We will see another example of
io.StringIO use a bit further on (➤ 227).
Command-Line Programming
If we need a program to be able to process text that may have been redirected
in the console or that may be in files listed on the command line, we can use
the fileinput module’s fileinput.input() function. This function iterates over
all the lines redirected from the console (if any) and over all the lines in the
files listed on the command line, as one continuous sequence of lines. The
module can report the current filename and line number at any time using
fileinput.filename() and fileinput.lineno(), and can handle some kinds of
compressed files.
Two separate modules are provided for handling command-line options,
optparse and getopt. The getopt module is popular because it is simple to use
Overview of Python’s Standard Library
and has been in the library for a long time. The optparse module is newer and
more powerful.
Example: The optparse Module
97 ➤
Back in Chapter 2 we described the program. In that chapter’s exercises we proposed extending the program to accept the command-line arguments, “maxwidth” taking an integer and “format” taking a string. The model solution ( has a 26-line function to process the arguments.
Here is the start of the main() function for, a version of the
program that uses the optparse module to handle the command-line arguments
rather than a custom function:
def main():
parser = optparse.OptionParser()
parser.add_option("-w", "--maxwidth", dest="maxwidth", type="int",
help=("the maximum number of characters that can be "
"output to string fields [default: %default]"))
parser.add_option("-f", "--format", dest="format",
help=("the format used for outputting numbers "
"[default: %default]"))
parser.set_defaults(maxwidth=100, format=".0f")
opts, args = parser.parse_args()
Only nine lines of code are needed, plus the import optparse statement. Furthermore, we do not need to explicitly provide -h and --help options; these are
handled by the optparse module to produce a suitable usage message using the
texts from the help keyword arguments, and with any “%default” text replaced
with the option’s default value.
Notice also that the options now use the conventional Unix style of having both
short and long option names that start with a hyphen. Short names are convenient for interactive use at the console; long names are more understandable
when used in shell scripts. For example, to set the maximum width to 80 we
can use any of -w80, -w 80, --maxwidth=80, or --maxwidth 80. After the command
line is parsed, the options are available using the dest names, for example,
opts.maxwidth and opts.format. Any command-line arguments that have not
been processed (usually filenames) are in the args list.
If an error occurs when parsing the command line, the optparse parser will
call sys.exit(2). This leads to a clean program termination and returns 2 to
the operating system as the program’s result value. Conventionally, a return
value of 2 signifies a usage error, 1 signifies any other kind of error, and 0
means success. When sys.exit() is called with no arguments it returns 0 to the
operating system.
Chapter 5. Modules
Mathematics and Numbers
In addition to the built-in int, float, and complex numbers, the library provides
the decimal.Decimal and fractions.Fraction numbers. Three numeric libraries
are available: math for the standard mathematical functions, cmath for complex
number mathematical functions, and random which provides many functions for
random number generation; these modules were introduced in Chapter 2.
Python’s numeric abstract base classes (classes that can be inherited from
but that cannot be used directly) are in the numbers module. They are useful
for checking that an object, say, x, is any kind of number using isinstance(x,
numbers.Number), or is a specific kind of number, for example, isinstance(x,
numbers.Rational) or isinstance(x, numbers.Integral).
Those involved in scientific and engineering programming will find the thirdparty NumPy package to be useful. This module provides highly efficient n-dimensional arrays, basic linear algebra functions and Fourier transforms, and
tools for integration with C, C++, and Fortran code. The SciPy package incorporates NumPy and extends it to include modules for statistical computations,
signal and image processing, genetic algorithms, and a great deal more. Both
are freely available from
Times and Dates
The calendar and datetime modules provide functions and classes for date and
time handling. However, they are based on an idealized Gregorian calendar,
so they are not suitable for dealing with pre-Gregorian dates. Time and date
handling is a very complex topic—the calendars in use have varied in different places and at different times, a day is not precisely 24 hours, a year is not
exactly 365 days, and daylight saving time and time zones vary. The datetime.datetime class (but not the class) has provisions for handling time zones, but does not do so out of the box. Third-party modules are
available to make good this deficiency, for example, dateutil from www.labix.
org/python-dateutil, and mxDateTime from
The time module handles timestamps. These are simply numbers that hold the
number of seconds since the epoch (1970-01-01T00:00:00 on Unix). This module can be used to get a timestamp of the machine’s current time in UTC (Coordinated Universal Time), or as a local time that accounts for daylight saving
time, and to create date, time, and date/time strings formatted in various ways.
It can also parse strings that have dates and times.
Overview of Python’s Standard Library
Example: The calendar, datetime, and time Modules
Objects of type datetime.datetime are usually created programmatically,
whereas objects that hold UTC date/times are usually received from external
sources, such as file timestamps. Here are some examples:
import calendar, datetime, time
moon_datetime_a = datetime.datetime(1969, 7, 20, 20, 17, 40)
moon_time = calendar.timegm(moon_datetime_a.utctimetuple())
moon_datetime_b = datetime.datetime.utcfromtimestamp(moon_time)
# returns: '1969-07-20T20:17:40'
# returns: '1969-07-20T20:17:40'
time.strftime("%Y-%m-%dT%H:%M:%S", time.gmtime(moon_time))
The moon_datetime_a variable is of type datetime.datetime and holds the
date and time that Apollo 11 landed on the moon. The moon_time variable
is of type int and holds the number of seconds since the epoch to the moon
landing—this number is provided by the calendar.timegm() function which
takes a time_struct object returned by the datetime.datetime.utctimetuple()
function, and returns the number of seconds that the time_struct represents.
(Since the moon landing occurred before the Unix epoch, the number is negative.) The moon_datetime_b variable is of type datetime.datetime and is created
from the moon_time integer to show the conversion from the number of seconds
since the epoch to a datetime.datetime object.★ The last three lines all return
identical ISO 8601-format date/time strings.
The current UTC date/time is available as a datetime.datetime object by calling
datetime.datetime.utcnow(), and as the number of seconds since the epoch by
calling time.time(). For the local date/time, use or
Algorithms and Collection Data Types
The bisect module provides functions for searching sorted sequences such
as sorted lists, and for inserting items while preserving the sort order. This
module’s functions use the binary search algorithm, so they are very fast. The
heapq module provides functions for turning a sequence such as a list into a
heap—a collection data type where the first item (at index position 0) is always
the smallest item, and for inserting and removing items while keeping the
sequence as a heap.
Unfortunately for Windows users, the datetime.datetime.utcfromtimestamp() function can’t handle
negative timestamps, that is, timestamps for dates prior to January 1, 1970.
135 ➤
111 ➤
136 ➤
Chapter 5. Modules
The collections package provides the collections.defaultdict dictionary and
the collections.namedtuple collection data types that we have previously discussed. In addition, this package provides the collections.UserList and collections.UserDict types, although subclassing the built-in list and dict types
is probably more common than using these types. Another type is collections.deque, which is similar to a list, but whereas a list is very fast for adding
and removing items at the end, a collections.deque is very fast for adding and
removing items both at the beginning and at the end.
Python 3.1 introduced the collections.OrderedDict and the collections.Counter
classes. OrderedDicts have the same API as normal dicts, although when
iterated the items are always returned in insertion order (i.e., from first to last
inserted), and the popitem() method always returns the most recently added
(i.e., last) item. The Counter class is a dict subclass used to provide a fast and
easy way of keeping various counts. Given an iterable or a mapping (such as
a dictionary), a Counter instance can, for example, return a list of the unique
elements or a list of the most common elements as (element, count) 2-tuples.
Python’s non-numeric abstract base classes (classes that can be inherited from
but that cannot be used directly) are also in the collections package. They are
discussed in Chapter 8.
The array module provides the array.array sequence type that can store numbers or characters in a very space-efficient way. It has similar behavior to lists
except that the type of object it can store is fixed when it is created, so unlike
lists it cannot store objects of different types. The third-party NumPy package
mentioned earlier also provides efficient arrays.
The weakref module provides functionality for creating weak references—these
behave like normal object references, except that if the only reference to an object is a weak reference, the object can still be scheduled for garbage collection.
This prevents objects from being kept in memory simply because we have a reference to them. Naturally, we can check whether the object a weak reference
refers to still exists, and can access the object if it does.
Example: The heapq Module
The heapq module provides functions for converting a list into a heap and for
adding and removing items from the heap while preserving the heap property.
A heap is a binary tree that respects the heap property, which is that the
first item (at index position 0) is always the smallest item.★ Each of a heap’s
subtrees is also a heap, so they too respect the heap property. Here is how a
heap could be created from scratch:
Strictly speaking, the heapq module provides a min heap; heaps where the first item is always the
largest are max heaps.
Overview of Python’s Standard Library
import heapq
heap = []
heapq.heappush(heap, (5, "rest"))
heapq.heappush(heap, (2, "work"))
heapq.heappush(heap, (4, "study"))
If we already have a list, we can turn it into a heap with heapq.heapify(alist);
this will do any necessary reordering in-place. The smallest item can be
removed from the heap using heapq.heappop(heap).
for x in heapq.merge([1, 3, 5, 8], [2, 4, 7], [0, 1, 6, 8, 9]):
print(x, end=" ") # prints: 0 1 1 2 3 4 5 6 7 8 8 9
The heapq.merge() function takes any number of sorted iterables as arguments
and returns an iterator that iterates over all the items from all the iterables
in order.
File Formats, Encodings, and Data Persistence
91 ➤
The standard library has extensive support for a variety of standard file formats and encodings. The base64 module has functions for reading and writing
using the Base16, Base32, and Base64 encodings specified in RFC 3548.★ The
quopri module has functions for reading and writing “quoted-printable” format. This format is defined in RFC 1521 and is used for MIME (Multipurpose
Internet Mail Extensions) data. The uu module has functions for reading and
writing uuencoded data. RFC 1832 defines the External Data Representation
Standard and module xdrlib provides functions for reading and writing data
in this format.
Modules are also provided for reading and writing archive files in the most
popular formats. The bz2 module can handle .bz2 files, the gzip module handles
.gz files, the tarfile module handles .tar, .tar.gz (also .tgz), and .tar.bz2 files,
and the zipfile module handles .zip files. We will see an example of using the
tarfile module in this subsection, and later on (➤ 227) there is a small example
that uses the gzip module; we will also see the gzip module in action again in
Chapter 7.
Support is also provided for handling some audio formats, with the aifc module for AIFF (Audio Interchange File Format) and the wave module for (uncompressed) .wav files. Some forms of audio data can be manipulated using the
audioop module, and the sndhdr module provides a couple of functions for determining what kind of sound data is stored in a file and some of its properties,
such as the sampling rate.
RFC (Request for Comments) documents are used to specify various Internet technologies.
Each one has a unique identification number and many of them have become officially adopted
Chapter 5. Modules
A format for configuration files (similar to old-style Windows .ini files) is
specified in RFC 822, and the configparser module provides functions for
reading and writing such files.
Many applications, for example, Excel, can read and write CSV (Comma
Separated Value) data, or variants such as tab-delimited data. The csv module
can read and write these formats, and can account for the idiosyncracies that
prevent CSV files from being straightforward to handle directly.
In addition to its support of various file formats, the standard library also has
packages and modules that provide data persistence. The pickle module is
used to store and retrieve arbitrary Python objects (including entire collections) to and from disk; this module is covered in Chapter 7. The library also
supports DBM files of various kinds—these are like dictionaries except that
their items are stored on disk rather than in memory, and both their keys and
their values must be bytes objects or strings. The shelve module, covered in
Chapter 12, can be used to provide DBM files with string keys and arbitrary
Python objects as values—the module seamlessly converts the Python objects to and from bytes objects behind the scenes. The DBM modules, Python’s
database API, and using the built-in SQLite database are all covered in Chapter 12.
Example: The base64 Module
The base64 module is mostly used for handling binary data that is embedded in
emails as ASCII text. It can also be used to store binary data inside .py files.
The first step is to get the binary data into Base64 format. Here we assume
that the base64 module has been imported and that the path and filename of a
.png file are in the variable left_align_png:
binary = open(left_align_png, "rb").read()
ascii_text = ""
for i, c in enumerate(base64.b64encode(binary)):
if i and i % 68 == 0:
ascii_text += "\\\n"
ascii_text += chr(c)
This code snippet reads the file in binary mode and converts it to a Base64
string of ASCII characters. Every sixty-eighth character a backslash-newline
combination is added. This limits the width of the lines of ASCII characters
to 68, but ensures that when the data is read back the newlines will be ignored
(because the backslash will escape them). The ASCII text obtained like this can
be stored as a bytes literal in a .py file, for example:
➤ 293
Overview of Python’s Standard Library
We’ve omitted most of the lines as indicated by the ellipsis.
The data can be converted back to its original binary form like this:
binary = base64.b64decode(LEFT_ALIGN_PNG)
The binary data could be written to a file using open(filename, "wb").write(
binary). Keeping binary data in .py files is much less compact than keeping
it in its original form, but can be useful if we want to provide a program that
requires some binary data as a single .py file.
Example: The tarfile Module
Most versions of Windows don’t come with support for the .tar format that
is so widely used on Unix systems. This inconvenient omission can easily be
rectified using Python’s tarfile module, which can create and unpack .tar and
.tar.gz archives (known as tarballs), and with the right libraries installed,
.tar.bz2 archives. The program can unpack tarballs using the tarfile
module; here we will just show some key extracts, starting with the first import
import bz2
except ImportError:
The bz2 module is used to handle the bzip2 compression format, but importing
it will fail if Python was built without access to the bzip2 library. (The Python
binary for Windows is always built with bzip2 compression built-in; it is only
on some Unix builds that it might be absent.) We account for the possibility
that the module is not available using a try … except block, and keep a Boolean
variable that we can refer to later (although we don’t quote the code that
uses it).
UNTRUSTED_PREFIXES = tuple(["/", "\\"] +
[c + ":" for c in string.ascii_letters])
This statement creates the tuple ('/', '\', 'A:', 'B:', …, 'Z:', 'a:', 'b:',
…, 'z:'). Any filename in the tarball being unpacked that begins with one of
these is suspect—tarballs should not use absolute paths since then they risk
overwriting system files, so as a precaution we will not unpack any file whose
name starts with one of these prefixes.
Chapter 5. Modules
def untar(archive):
tar = None
tar =
for member in tar.getmembers():
print("untrusted prefix, ignoring",
elif ".." in
print("suspect path, ignoring",
except (tarfile.TarError, EnvironmentError) as err:
if tar is not None:
Each file in a tarball is called a member. The tarfile.getmembers() function
returns a list of tarfile.TarInfo objects, one for each member. The member’s
filename, including its path, is in the attribute. If the
name begins with an untrusted prefix, or contains .. in its path, we output an
error message; otherwise, we call tarfile.extract() to save the member to disk.
The tarfile module has its own set of custom exceptions, but we have taken the
simplistic approach that if any exception occurs we output the error message
and finish.
def error(message, exit_status=1):
We have just quoted the error() function for completeness. The (unquoted)
main() function prints a usage message if -h or --help is given; otherwise, it
performs some basic checks before calling untar() with the tarball’s filename.
File, Directory, and Process Handling
The shutil module provides high-level functions for file and directory handling,
including shutil.copy() and shutil.copytree() for copying files and entire
directory trees, shutil.move() for moving directory trees, and shutil.rmtree()
for removing entire directory trees, including nonempty ones.
Temporary files and directories should be created using the tempfile module
which provides the necessary functions, for example, tempfile.mkstemp(), and
creates the temporaries in the most secure manner possible.
Overview of Python’s Standard Library
The filecmp module can be used to compare files with the filecmp.cmp() function and to compare entire directories with the filecmp.cmpfiles() function.
One very powerful and effective use of Python programs is to orchestrate
the running of other programs. This can be done using the subprocess module which can start other processes, communicate with them using pipes, and
retrieve their results. This module is covered in Chapter 10. An even more
powerful alternative is to use the multiprocessing module which provides extensive facilities for offloading work to multiple processes and for accumulating
results, and can often be used as an alternative to multithreading.
The os module provides platform-independent access to operating system functionality. The os.environ variable holds a mapping object whose items are environment variable names and their values. The program’s working directory
is provided by os.getcwd() and can be changed using os.chdir(). The module
also provides functions for low-level file-descriptor-based file handling. The
os.access() function can be used to determine whether a file exists or whether
it is readable or writable, and the os.listdir() function returns a list of the
entries (e.g., the files and directories, but excluding the . and .. entries), in the
directory it is given. The os.stat() function returns various items of information about a file or directory, such as its mode, access time, and size.
Directories can be created using os.mkdir(), or if intermediate directories
need to be created, using os.makedirs(). Empty directories can be removed
using os.rmdir(), and directory trees that contain only empty directories can
be removed using os.removedirs(). Files or directories can be removed using
os.remove(), and can be renamed using os.rename().
The os.walk() function iterates over an entire directory tree, retrieving the
name of every file and directory in turn.
The os module also provides many low-level platform-specific functions, for
example, to work with file descriptors, and to fork (only on Unix systems),
spawn, and exec.
Whereas the os module provides functions for interacting with the operating
system, especially in the context of the file system, the os.path module provides a mixture of string manipulation (of paths), and some file system convenience functions. The os.path.abspath() function returns the absolute path
of its argument, with redundant path separators and .. elements removed.
The os.path.split() function returns a 2-tuple with the first element containing the path and the second the filename (which will be empty if a path
with no filename was given). These two parts are also available directly using
os.path.basename() and os.path.dirname(). A filename can also be split into
two parts, name and extension, using os.path.splitext(). The os.path.join()
function takes any number of path strings and returns a single path using the
platform-specific path separator.
Chapter 5. Modules
If we need several pieces of information about a file or directory we can use
os.stat(), but if we need just one piece, we can use the relevant os.path
function, for example, os.path.exists(), os.path.getsize(), os.path.isfile(), or
The mimetypes module has the mimetypes.guess_type() function that tries to
guess the given file’s MIME type.
Example: The os and os.path Modules
Here is how we can use the os and os.path modules to create a dictionary
where each key is a filename (including its path) and where each value is the
timestamp (seconds since the epoch) when the file was last modified, for those
files in the given path:
date_from_name = {}
for name in os.listdir(path):
fullname = os.path.join(path, name)
if os.path.isfile(fullname):
date_from_name[fullname] = os.path.getmtime(fullname)
This code is pretty straightforward, but can be used only for the files in a
single directory. If we need to traverse an entire directory tree we can use the
os.walk() function.
Here is a code snippet taken from the program.★ The code creates a
dictionary where each key is a 2-tuple (file size, filename) where the filename
excludes the path, and where each value is a list of the full filenames that
match their key’s filename and have the same file size:
data = collections.defaultdict(list)
for root, dirs, files in os.walk(path):
for filename in files:
fullname = os.path.join(root, filename)
key = (os.path.getsize(fullname), filename)
For each directory, os.walk() returns the root and two lists, one of the subdirectories in the directory and the other of the files in the directory. To get the full
path for a filename we need to combine just the root and the filename. Notice
that we do not have to recurse into the subdirectories ourselves—os.walk() does
that for us. Once the data has been gathered, we can iterate over it to produce
a report of possible duplicate files:
A much more sophisticated find duplicates program,, which uses multiple
threads and MD5 checksums, is covered in Chapter 10.
Overview of Python’s Standard Library
for size, filename in sorted(data):
names = data[(size, filename)]
if len(names) > 1:
print("{filename} ({size} bytes) may be duplicated "
"({0} files):".format(len(names), **locals()))
for name in names:
Because the dictionary keys are (size, filename) tuples, we don’t need to use a
key function to get the data sorted in size order. If any (size, filename) tuple
has more than one filename in its list, these might be duplicates.
shell32.dll (8460288 bytes) may be duplicated (2 files):
This is the last item taken from the 3 282 lines of output produced by running \windows on a Windows XP system.
Networking and Internet Programming
Packages and modules for networking and Internet programming are a major
part of Python’s standard library. At the lowest level, the socket module provides the most fundamental network functionality, with functions for creating
sockets, doing DNS (Domain Name System) lookups, and handling IP (Internet
Protocol) addresses. Encrypted and authenticated sockets can be set up using
the ssl module. The socketserver module provides TCP (Transmission Control
Protocol) and UDP (User Datagram Protocol) servers. These servers can handle requests directly, or can create a separate process (by forking) or a separate
thread to handle each request. Asynchronous client and server socket handling can be achieved using the asyncore module and the higher-level asynchat
module that is built on top of it.
Python has defined the WSGI (Web Server Gateway Interface) to provide
a standard interface between web servers and web applications written in
Python. In support of the standard the wsgiref package provides a reference
implementation of WSGI that has modules for providing WSGI-compliant
HTTP servers, and for handling response header and CGI (Common Gateway
Interface) scripts. In addition, the http.server module provides an HTTP server which can be given a request handler (a standard one is provided), to run
CGI scripts. The http.cookies and http.cookiejar modules provide functions
for managing cookies, and CGI script support is provided by the cgi and cgitb
Chapter 5. Modules
Client access to HTTP requests is provided by the http.client module, although
the higher-level urllib package’s modules, urllib.parse, urllib.request, urllib.response, urllib.error, and urllib.robotparser, provide easier and more
convenient access to URLs. Grabbing a file from the Internet is as simple as:
fh = urllib.request.urlopen("")
html ="utf8")
The urllib.request.urlopen() function returns an object that behaves much
like a file object opened in read binary mode. Here we retrieve the Python
Web site’s index.html file (as a bytes object), and store it as a string in the html
variable. It is also possible to grab files and store them in local files with the
urllib.request.urlretrieve() function.
HTML and XHTML documents can be parsed using the html.parser module,
URLs can be parsed and created using the urllib.parse module, and robots.txt
files can be parsed with the urllib.robotparser module. Data that is represented using JSON (JavaScript Object Notation) can be read and written using the
json module.
In addition to HTTP server and client support, the library provides XML-RPC
(Remote Procedure Call) support with the xmlrpc.client and xmlrpc.server
modules. Additional client functionality is provided for FTP (File Transfer
Protocol) by the ftplib module, for NNTP (Network News Transfer Protocol)
by the nntplib module, and for TELNET with the telnetlib module.
The smtpd module provides an SMTP (Simple Mail Transfer Protocol) server,
and the email client modules are smtplib for SMTP, imaplib for IMAP4 (Internet Message Access Protocol), and poplib for POP3 (Post Office Protocol). Mailboxes in various formats can be accessed using the mailbox module. Individual
messages (including multipart messages) can be created and manipulated using the email module.
If the standard library’s packages and modules are insufficient in this
area, Twisted ( provides a comprehensive third-party networking library. Many third-party web programming libraries are
also available, including Django ( and Turbogears
( for creating web applications, and Plone (
and Zope ( which provide complete web frameworks and content
management systems. All of these libraries are written in Python.
There are two widely used approaches to parsing XML documents. One is the
DOM (Document Object Model) and the other is SAX (Simple API for XML).
Two DOM parsers are provided, one by the xml.dom module and the other by
the xml.dom.minidom module. A SAX parser is provided by the xml.sax
Overview of Python’s Standard Library
ule. We have already used the xml.sax.saxutils module for its xml.sax.saxutils.escape() function (to XML-escape “&”, “<”, and “>”). There is also an
xml.sax.saxutils.quoteattr() function that does the same thing but additionally escapes quotes (to make the text suitable for a tag’s attribute), and
xml.sax.saxutils.unescape() to do the opposite conversion.
Two other parsers are available. The xml.parsers.expat module can be used to
parse XML documents with expat, providing the expat library is available, and
the xml.etree.ElementTree can be used to parse XML documents using a kind
of dictionary/list interface. (By default, the DOM and element tree parsers
themselves use the expat parser under the hood.)
Writing XML manually and writing XML using DOM and element trees, and
parsing XML using the DOM, SAX, and element tree parsers, is covered in
Chapter 7.
There is also a third-party library, lxml (, that claims
to be “the most feature-rich and easy-to-use library for working with XML
and HTML in the Python language.” This library provides an interface that
is essentially a superset of what the element tree module provides, as well as
many additional features such as support for XPath, XSLT, and many other
XML technologies.
Example: The xml.etree.ElementTree Module
Python’s DOM and SAX parsers provide the APIs that experienced XML
programmers are used to, and the xml.etree.ElementTree module offers a more
Pythonic approach to parsing and writing XML. The element tree module is
a fairly recent addition to the standard library,★ and so may not be familiar to
some readers. In view of this, we will present a very short example here to give
a flavor of it—Chapter 7 provides a more substantial example and provides
comparative code using DOM and SAX.
The U.S. government’s NOAA (National Oceanic and Atmospheric Administration) Web site provides a wide variety of data, including an XML file that lists
the U.S. weather stations. The file is more than 20 000 lines long and contains
details of around two thousand stations. Here is a typical entry:
<station_name>Boston, Logan International Airport</station_name>
The xml.etree.ElementTree module first appeared in Python 2.5.
Chapter 5. Modules
We have cut out a few lines and reduced the indentation that is present in the
file. The file is about 840K in size, so we have compressed it using gzip to a
more manageable 72K. Unfortunately, the element tree parser requires either
a filename or a file object to read, but we cannot give it the compressed file since
that will just appear to be random binary data. We can solve this problem with
two initial steps:
binary =
fh = io.StringIO(binary.decode("utf8"))
213 ➤
The gzip module’s function is similar to the built-in open() except
that it reads gzip-compressed files (those with extension .gz) as raw binary
data. We need the data available as a file that the element tree parser can
work with, so we use the bytes.decode() method to convert the binary data to a
string using UTF-8 encoding (which is what the XML file uses), and we create
a file-like io.StringIO object with the string containing the entire XML file as
its data.
tree = xml.etree.ElementTree.ElementTree()
root = tree.parse(fh)
stations = []
for element in tree.getiterator("station_name"):
Here we create a new xml.etree.ElementTree.ElementTree object and give it a file
object from which to read the XML we want it to parse. As far as the element
tree parser is concerned it has been passed a file object open for reading,
although in fact it is reading a string inside an io.StringIO object. We want to
extract the names of all the weather stations, and this is easily achieved using
the xml.etree.ElementTree.ElementTree.getiterator() method which returns an
iterator that returns all the xml.etree.ElementTree.Element objects that have
the given tag name. We just use the element’s text attribute to retrieve the
text. Like os.walk(), we don’t have to do any recursion ourselves; the iterator
method does that for us. Nor do we have to specify a tag—in which case the
iterator will return every element in the entire XML document.
Other Modules
We don’t have the space to cover the nearly 200 packages and modules that are
available in the standard library. Nonetheless, this general overview should
be sufficient to get a flavor of what the library provides and some of the key
packages in the major areas it serves. In this section’s final subsection we
discuss just a few more areas of interest.
In the previous section we saw how easy it is to create tests in docstrings and
to run them using the doctest module. The library also has a unit-testing
➤ 293
Overview of Python’s Standard Library
framework provided by the unittest module—this is a Python version of the
Java JUnit test framework. The doctest module also provides some basic integration with the unittest module. (Testing is covered more fully in Chapter 9.) Several third-party testing frameworks are also available, for example,
py.test from and nose from
Noninteractive applications such as servers often report problems by writing
to log files. The logging module provides a uniform interface for logging, and
in addition to being able to log to files, it can log using HTTP GET or POST
requests, or using email or sockets.
The library provides many modules for introspection and code manipulation,
and although most of them are beyond the scope of this book, one that is worth
mentioning is pprint which has functions for “pretty printing” Python objects,
including collection data types, which is sometimes useful for debugging. We
will see a simple use of the inspect module that introspects live objects in
Chapter 8.
The threading module provides support for creating threaded applications,
and the queue module provides three different kinds of thread-safe queues.
Threading is covered in Chapter 10.
Python has no native support for GUI programming, but several GUI libraries
can be used by Python programs. The Tk library is available using the tkinter
module, and is usually installed as standard. GUI programming is introduced
in Chapter 15.
The abc (Abstract Base Class) module provides the functions necessary for
creating abstract base classes. This module is covered in Chapter 8.
146 ➤
The copy module provides the copy.copy() and copy.deepcopy() functions that
were discussed in Chapter 3.
Access to foreign functions, that is, to functions in shared libraries (.dll files on
Windows, .dylib files on Mac OS X, and .so files on Linux), is available using
the ctypes module. Python also provides a C API, so it is possible to create
custom data types and functions in C and make these available to Python.
Both the ctypes module and Python’s C API are beyond the scope of this book.
If none of the packages and modules mentioned in this section provides
the functionality you need, before writing anything from scratch it is worth
checking the Python documentation’s Global Module Index to see whether
a suitable module is available, since we have not been able to mention every one here. And failing that, try looking at the Python Package Index
( which contains several thousand Python add-ons ranging from small one-file modules all the way up to large library and framework
packages containing anything from scores to hundreds of modules.
Chapter 5. Modules
The chapter began by introducing the various syntaxes that can be used for
importing packages, modules, and objects inside modules. We noted that
many programmers only use the import importable syntax so as to avoid name
clashes, and that we must be careful not to give a program or module the same
name as a top-level Python module or directory.
Also discussed were Python packages. These are simply directories with an file and one or more .py modules inside them. The
file can be empty, but to support the from importable import * syntax, we can
create an __all__ special variable in the file set to a list of module
names. We can also put any common initialization code in the file.
It was noted that packages can be nested simply by creating subdirectories and
having each of these contain its own file.
Two custom modules were described. The first just provided a few functions
and had very simple doctests. The second was more elaborate with its own
exceptions, the use of dynamic function creation to create a function with a
platform-specific implementation, private global data, a call to an initialization
function, and more elaborate doctests.
About half the chapter was devoted to a high-level overview of Python’s standard library. Several string handling modules were mentioned and a couple
of io.StringIO examples were presented. One example showed how to write
text to a file using either the built-in print() function or a file object’s write()
method, and how to use an io.StringIO object in place of a real file. In previous
chapters we handled command-line options by reading sys.argv ourselves, but
in the coverage of the library’s support for command-line programming we introduced the optparse module which greatly simplifies command-line argument
handling—we will use this module extensively from now on.
Mention was made of Python’s excellent support for numbers, and the library’s
numeric types and its three modules of mathematical functions, as well as
the support for scientific and engineering mathematics provided by the SciPy
project. Both library and third-party date/time handling classes were briefly
described and examples of how to obtain the current date/time and how to
convert between datetime.datetime and the number of seconds since the epoch
were shown. Also discussed were the additional collection data types and the
algorithms for working with ordered sequences that the standard library
provides, along with some examples of using the heapq module’s functions.
The modules that support various file encodings (besides character encodings)
were discussed, as well as the modules for packing and unpacking the most
popular archive formats, and those that have support for audio data. An example showing how to use the Base64 encoding to store binary data in .py files was
given, and also a program to unpack tarballs. Considerable support is provided
for handling directories and files—and all of this is abstracted into platformindependent functions. Examples were shown for creating a dictionary with
filename keys and last modified timestamp values, and for doing a recursive
search of a directory to identify possible duplicate files based on their name
and size.
A large part of the library is devoted to networking and Internet programming.
We very briefly surveyed what is available, from raw sockets (including
encrypted sockets), to TCP and UDP servers, to HTTP servers and support for
the WSGI. Also mentioned were the modules for handling cookies, CGI scripts,
and HTTP data, and for parsing HTML, XHTML, and URLs. Other modules
that were mentioned included those for handling XML-RPC and for handling
higher-level protocols such as FTP and NNTP, as well as the email client and
server support using SMTP and client support for IMAP4 and POP3.
The library’s comprehensive support for XML writing and parsing was also
mentioned, including the DOM, SAX, and element tree parsers, and the expat
module. And an example was given using the element tree module. Mention
was also made of some of the many other packages and modules that the
library provides.
Python’s standard library represents an extremely useful resource that can
save enormous amounts of time and effort, and in many cases allows us to
write much smaller programs by relying on the functionality that the library
provides. In addition, literally thousands of third-party packages are available
to fill any gaps the standard library may have. All of this predefined functionality allows us to focus much more on what we want our programs to do, while
leaving the library modules to take care of most of the details.
This chapter brings us to the end of the fundamentals of procedural programming. Later chapters, and particularly Chapter 8, will look at more advanced
and specialized procedural techniques, and the following chapter introduces
object-oriented programming. Using Python as a purely procedural language is
both possible and practical—especially for small programs—but for medium to
large programs, for custom packages and modules, and for long-term maintainability, the object-oriented approach usually wins out. Fortunately, all that we
have covered up to now is both useful and relevant in object-oriented programming, so the subsequent chapters will continue to build up our Python knowledge and skills based on the foundations that have now been laid.
Write a program to show directory listings, rather like the dir command in
Windows or ls in Unix. The benefit of creating our own listing program is
that we can build in the defaults we prefer and can use the same program on
Chapter 5. Modules
all platforms without having to remember the differences between dir and ls.
Create a program that supports the following interface:
Usage: [options] [path1 [path2 [... pathN]]]
The paths are optional; if not given . is used.
-h, --help
show this help message and exit
-H, --hidden
show hidden files [default: off]
-m, --modified show last modified date/time [default: off]
-o ORDER, --order=ORDER
order by ('name', 'n', 'modified', 'm', 'size', 's') [default: name]
-r, --recursive recurse into subdirectories [default: off]
-s, --sizes
show sizes [default: off]
(The output has been modified slightly to fit the book’s page.)
Here is an example of output on a small directory using the command line -ms -os misc/:
2008-02-11 14:17:03
2008-02-05 14:22:38
2007-12-13 12:01:14
12,184 misc/abstract.pdf
109,788 misc/klmqtintro.lyx
1,359,950 misc/tracking.pdf
3 files, 1 directory
We used option grouping in the command line (optparse handles this automatically for us), but the same could have been achieved using separate options, for
example, -m -s -os misc/, or by even more grouping, -msos misc/, or
by using long options, --modified --sizes --order=size misc/, or any combination of these. Note that we define a “hidden” file or directory as one whose
name begins with a dot (.).
The exercise is quite challenging. You will need to read the optparse documentation to see how to provide options that set a True value, and how to offer a
fixed list of choices. If the user sets the recursive option you will need to process the files (but not the directories) using os.walk(); otherwise, you will have
to use os.listdir() and process both files and directories yourself.
One rather tricky aspect is avoiding hidden directories when recursing. They
can be cut out of os.walk()’s dirs list—and therefore skipped by os.walk()—by
modifying that list. But be careful not to assign to the dirs variable itself, since
that won’t change the list it refers to but will simply (and uselessly) replace it;
the approach used in the model solution is to assign to a slice of the whole list,
that is, dirs[:] = [dir for dir in dirs if not dir.startswith(".")].
86 ➤
The best way to get grouping characters in the file sizes is to import the locale
module, call locale.setlocale() to get the user’s default locale, and use the n
format character. Overall, is about 130 lines split over four functions.
● The Object-Oriented Approach
● Custom Classes
● Custom Collection Classes
In all the previous chapters we used objects extensively, but our style of
programming has been strictly procedural. Python is a multiparadigm
language—it allows us to program in procedural, object-oriented, and functional style, or in any mixture of styles, since it does not force us to program in any
one particular way.
It is perfectly possible to write any program in procedural style, and for very
small programs (up to, say, 500 lines), doing so is rarely a problem. But for most
programs, and especially for medium-size and large programs, object-oriented
programming offers many advantages.
This chapter covers all the fundamental concepts and techniques for doing
object-oriented programming in Python. The first section is especially for those
who are less experienced and for those coming from a procedural programming
background (such as C or Fortran). The section starts by looking at some of
the problems that can arise with procedural programming that object-oriented
programming can solve. Then it briefly describes Python’s approach to objectoriented programming and explains the relevant terminology. After that, the
chapter’s two main sections begin.
The second section covers the creation of custom data types that hold single items (although the items themselves may have many attributes), and
the third section covers the creation of custom collection data types that can
hold any number of objects of any types. These sections cover most aspects
of object-oriented programming in Python, although we defer some more advanced material to Chapter 8.
Chapter 6. Object-Oriented Programming
The Object-Oriented Approach
In this section we will look at some of the problems of a purely procedural approach by considering a situation where we need to represent circles, potentially lots of them. The minimum data required to represent a circle is its (x, y)
position and its radius. One simple approach is to use a 3-tuple for each circle.
For example:
circle = (11, 60, 8)
One drawback of this approach is that it isn’t obvious what each element of
the tuple represents. We could mean (x, y, radius) or, just as easily, (radius, x, y). Another drawback is that we can access the elements by index
position only. If we have two functions, distance_from_origin(x, y) and
edge_distance_from_origin(x, y, radius), we would need to use tuple unpacking
to call them with a circle tuple:
distance = distance_from_origin(*circle[:2])
distance = edge_distance_from_origin(*circle)
Both of these assume that the circle tuples are of the form (x, y, radius).
We can solve the problem of knowing the element order and of using tuple
unpacking by using a named tuple:
import collections
Circle = collections.namedtuple("Circle", "x y radius")
circle = Circle(13, 84, 9)
distance = distance_from_origin(circle.x, circle.y)
This allows us to create Circle 3-tuples with named attributes which makes
function calls much easier to understand, since to access elements we can use
their names. Unfortunately, problems remain. For example, there is nothing
to stop an invalid circle from being created:
circle = Circle(33, 56, -5)
It doesn’t make sense to have a circle with a negative radius, but the circle
named tuple is created here without raising an exception—just as it would be
if the radius was given as a variable that held a negative number. The error
will be noticed only if we call the edge_distance_from_origin() function—and
then only if that function actually checks for a negative radius. This inability
to validate when creating an object is probably the worst aspect of taking a
purely procedural approach.
If we want circles to be mutable so that we can move them by changing their
coordinates or resize them by changing their radius, we can do so by using the
private collections.namedtuple._replace() method:
The Object-Oriented Approach
circle = circle._replace(radius=12)
Just as when we create a Circle, there is nothing to stop us from (or warn us
about) setting invalid data.
If the circles were going to need lots of changes, we might opt to use a mutable
data type such as a list, for the sake of convenience:
circle = [36, 77, 8]
This doesn’t give us any protection from putting in invalid data, and the best
we can do about accessing elements by name is to create some constants so that
we can write things like circle[RADIUS] = 5. But using a list brings additional
problems—for example, we can legitimately call circle.sort()! Using a dictionary might be an alternative, for example, circle = dict(x=36, y=77, radius=8),
but again there is no way to ensure a valid radius and no way to prevent inappropriate methods from being called.
Object-Oriented Concepts and Terminology
What we need is some way to package up the data that is needed to represent
a circle, and some way to restrict the methods that can be applied to the data
so that only valid operations are possible. Both of these things can be achieved
by creating a custom Circle data type. We will see how to create a Circle data
type in later in this section, but first we need to cover some preliminaries and
explain some terminology. Don’t worry if the terminology is unfamiliar at first;
it will become much clearer once we reach the examples.
We use the terms class, type, and data type interchangeably. In Python we
can create custom classes that are fully integrated and that can be used just
like the built-in data types. We have already encountered many classes, for
example, dict, int, and str. We use the term object, and occasionally the term
instance, to refer to an instance of a particular class. For example, 5 is an int
object and "oblong" is a str object.
Most classes encapsulate both data and the methods that can be applied to that
data. For example, the str class holds a string of Unicode characters as its data
and supports methods such as str.upper(). Many classes also support additional features; for example, we can concatenate two strings (or any two sequences)
using the + operator and find a sequence’s length using the built-in len() function. Such features are provided by special methods—these are like normal
methods except that their names always begin and end with two underscores,
and are predefined. For example, if we want to create a class that supports
concatenation using the + operator and also the len() function, we can do so by
implementing the __add__() and __len__() special methods in our class. Conversely, we should never define any method with a name that begins and ends
with two underscores unless it is one of the predefined special methods and is
Chapter 6. Object-Oriented Programming
appropriate to our class. This will ensure that we never get conflicts with later
versions of Python even if they introduce new predefined special methods.
Objects usually have attributes—methods are callable attributes, and other
attributes are data. For example, a complex object has imag and real attributes
and lots of methods, including special methods like __add__() and __sub__ (to
support the binary + and - operators), and normal methods like conjugate().
Data attributes (often referred to simply as “attributes”) are normally implemented as instance variables, that is, variables that are unique to a particular
object. We will see examples of this, and also examples of how to provide data
attributes as properties. A property is an item of object data that is accessed like
an instance variable but where the accesses are handled by methods behind the
scenes. As we will see, using properties makes it easy to do data validation.
Inside a method (which is just a function whose first argument is the instance
on which it is called to operate), several kinds of variables are potentially accessible. The object’s instance variables can be accessed by qualifying their name
with the instance itself. Local variables can be created inside the method; these
are accessed without qualification. Class variables (sometimes called static
variables) can be accessed by qualifying their name with the class name, and
global variables, that is, module variables, are accessed without qualification.
Some of the Python literature uses the concept of a namespace, a mapping from
names to objects. Modules are namespaces—for example, after the statement
import math we can access objects in the math module by qualifying them with
their namespace name (e.g., math.pi and math.sin()). Similarly, classes and objects are also namespaces; for example, if we have z = complex(1, 2), the z object’s namespace has two attributes which we can access (z.real and z.imag).
One of the advantages of object orientation is that if we have a class, we can
specialize it. This means that we make a new class that inherits all the attributes (data and methods) from the original class, usually so that we can add
or replace methods or add more instance variables. We can subclass (another
term for specialize), any Python class, whether built-in or from the standard
library, or one of our own custom classes.★ The ability to subclass is one of the
great advantages offered by object-oriented programming since it makes it
straightforward to use an existing class that has tried and tested functionality as the basis for a new class that extends the original, adding new data attributes or new functionality in a very clean and direct way. Furthermore, we
can pass objects of our new class to functions and methods that were written
for the original class and they will work correctly.
We use the term base class to refer to a class that is inherited; a base class
may be the immediate ancestor, or may be further up the inheritance tree.
Another term for base class is super class. We use the term subclass, derived
Some library classes that are implemented in C cannot be subclassed; such classes specify this in
their documentation.
Chapter 6. Object-Oriented Programming
Some object-oriented languages have two features that Python does not provide. The first is overloading, that is, having methods with the same name but
with different parameter lists in the same class. Thanks to Python’s versatile
argument-handling capabilities this is never a limitation in practice. The second is access control—there are no bulletproof mechanisms for enforcing data
privacy. However, if we create attributes (instance variables or methods) that
begin with two leading underscores, Python will prevent unintentional accesses so that they can be considered to be private. (This is done by name mangling;
we will see an example in Chapter 8.)
Just as we use an uppercase letter as the first letter of custom modules, we will
do the same thing for custom classes. We can define as many classes as we like,
either directly in a program or in modules—class names don’t have to match
module names, and modules may contain as many class definitions as we like.
Now that we have seen some of the problems that classes can solve, introduced
the necessary terminology, and covered some background matters, we can
begin to create some custom classes.
Custom Classes
In earlier chapters we created custom classes: custom exceptions. Here are two
new syntaxes for creating custom classes:
class className:
class className(base_classes):
Since the exception subclasses we created did not add any new attributes (no
instance data or methods) we used a suite of pass (i.e., nothing added), and
since the suite was just one statement we put it on the same line as the class
statement itself. Note that just like def statements, class is a statement, so
we can create classes dynamically if we want to. A class’s methods are created
using def statements in the class’s suite. Class instances are created by calling
the class with any necessary arguments; for example, x = complex(4, 8) creates
a complex number and sets x to be an object reference to it.
Attributes and Methods
Let’s start with a very simple class, Point, that holds an (x, y) coordinate. The
class is in file, and its complete implementation (excluding docstrings)
is show here:
class Point:
Custom Classes
def __init__(self, x=0, y=0):
self.x = x
self.y = y
def distance_from_origin(self):
return math.hypot(self.x, self.y)
def __eq__(self, other):
return self.x == other.x and self.y == other.y
def __repr__(self):
return "Point({0.x!r}, {0.y!r})".format(self)
def __str__(self):
return "({0.x!r}, {0.y!r})".format(self)
Since no base classes are specified, Point is a direct subclass of object, just
as though we had written class Point(object). Before we discuss each of the
methods, let’s see some examples of their use:
import Shape
a = Shape.Point()
b = Shape.Point(3, 4)
b.x = -19
a == b, a != b
# returns: 'Point(0, 0)'
# returns: '(3, 4)'
# returns: 5.0
# returns: '(-19, 4)'
# returns: (False, True)
The Point class has two data attributes, self.x and self.y, and five methods
(not counting inherited methods), four of which are special methods; they are
illustrated in Figure 6.2. Once the Shape module is imported, the Point class
can be used like any other. The data attributes can be accessed directly (e.g.,
y = a.y), and the class integrates nicely with all of Python’s other classes by
providing support for the equality operator (==) and for producing strings in
representational and string forms. And Python is smart enough to supply the
inequality operator (!=) based on the equality operator. (It is also possible to
specify each operator individually if we want total control, for example, if they
are not exact opposites of each other.)
Python automatically supplies the first argument in method calls—it is an
object reference to the object itself (called this in C++ and Java). We must include this argument in the parameter list, and by convention the parameter is
called self. All object attributes (data and method attributes) must be qualified
by self. This requires a little bit more typing compared with some other languages, but has the advantage of providing absolute clarity: we always know
that we are accessing an object attribute if we qualify with self.
Chapter 6. Object-Oriented Programming
Figure 6.2 The Point class’s inheritance hierarchy
To create an object, two steps are necessary. First a raw or uninitialized object
must be created, and then the object must be initialized, ready for use. Some
object-oriented languages (such as C++ and Java) combine these two steps
into one, but Python keeps them separate. When an object is created (e.g., p =
Shape.Point()), first the special method __new__() is called to create the object,
and then the special method __init__() is called to initialize it.
In practice almost every Python class we create will require us to reimplement only the __init__() method, since the object.__new__() method is almost always sufficient and is automatically called if we don’t provide our own
__new__() method. (Later in this chapter we will show a rare example where
we do need to reimplement __new__().) Not having to reimplement methods
in a subclass is another benefit of object-oriented programming—if the base
class method is sufficient we don’t have to reimplement it in our subclass.
This works because if we call a method on an object and the object’s class
does not have an implementation of that method, Python will automatically
go through the object’s base classes, and their base classes, and so on, until it
finds the method—and if the method is not found an AttributeError exception
is raised.
For example, if we execute p = Shape.Point(), Python begins by looking for
the method Point.__new__(). Since we have not reimplemented this method,
Python looks for the method in Point’s base classes. In this case there is only
one base class, object, and this has the required method, so Python calls object.__new__() and creates a raw uninitialized object. Then Python looks for
the initializer, __init__(), and since we have reimplemented it, Python doesn’t
need to look further and calls Point.__init__(). Finally, Python sets p to be an
object reference to the newly created and initialized object of type Point.
Because they are so short and a few pages away, for convenience we will show
each method again before discussing it.
➤ 256
Custom Classes
def __init__(self, x=0, y=0):
self.x = x
self.y = y
The two instance variables, self.x and self.y, are created in the initializer,
and assigned the values of the x and y parameters. Since Python will find this
initializer when we create a new Point object, the object.__init__() method
will not be called. This is because as soon as Python has found the required
method it calls it and doesn’t look further.
Object-oriented purists might start the method off with a call to the base
class __init__() method by calling super().__init__(). The effect of calling
the super() function like this is to call the base class’s __init__() method. For
classes that directly inherit object there is no need to do this, and in this book
we call base class methods only when necessary—for example, when creating
classes that are designed to be subclassed, or when creating classes that don’t
directly inherit object. This is to some extent a matter of coding style—it is
perfectly reasonable to always call super().__init__() at the start of a custom
class’s __init__() method.
def distance_from_origin(self):
return math.hypot(self.x, self.y)
This is a conventional method that performs a computation based on the
object’s instance variables. It is quite common for methods to be fairly short
and to have only the object they are called on as an argument, since often all
the data the method needs is available inside the object.
def __eq__(self, other):
return self.x == other.x and self.y == other.y
Methods should not have names that begin and end with two underscores—unless they are one of the predefined special methods. Python provides special methods for all the comparison operators as shown in Table 6.1.
All instances of custom classes support == by default, and the comparison
returns False—unless we compare a custom object with itself. We can override
this behavior by reimplementing the __eq__() special method as we have done
here. Python will supply the __ne__() (not equal) inequality operator (!=)
automatically if we implement __eq__() but don’t implement __ne__().
By default, all instances of custom classes are hashable, so hash() can be called
on them and they can be used as dictionary keys and stored in sets. But if we
reimplement __eq__(), instances are no longer hashable. We will see how to fix
this when we discuss the FuzzyBool class later on.
By implementing this special method we can compare Point objects, but if we
were to try to compare a Point with an object of a different type—say, int—we
➤ 254
Chapter 6. Object-Oriented Programming
Table 6.1 Comparison Special Methods
Special Method
Usage Description
__lt__(self, other)
x < y
__le__(self, other)
x <= y Returns True if x is less than or equal to y
__eq__(self, other)
x == y Returns True if x is equal to y
__ne__(self, other)
x != y Returns True if x is not equal to y
__ge__(self, other)
x >= y Returns True if x is greater than or equal to y
__gt__(self, other)
x > y
Returns True if x is less than y
Returns True if x is greater than y
would get an AttributeError exception (since ints don’t have an x attribute).
On the other hand, we can compare Point objects with other objects that
coincidentally just happen to have an x attribute (thanks to Python’s duck
typing), but this may lead to surprising results.
If we want to avoid inappropriate comparisons there are a few approaches
we can take. One is to use an assertion, for example, assert isinstance(other,
Point). Another is to raise a TypeError to indicate that comparisons between the
two types are not supported, for example, if not isinstance(other, Point): raise
TypeError(). The third way (which is also the most Pythonically correct) is to do
this: if not isinstance(other, Point): return NotImplemented. In this third case,
if NotImplemented is returned, Python will then try calling other.__eq__(self) to
see whether the other type supports the comparison with the Point type, and if
there is no such method or if that method also returns NotImplemented, Python
will give up and raise a TypeError exception. (Note that only reimplementations
of the comparison special methods listed in Table 6.1 may return NotImplemented.)
The built-in isinstance() function takes an object and a class (or a tuple of
classes), and returns True if the object is of the given class (or of one of the tuple
of classes), or of one of the class’s (or one of the tuple of classes’) base classes.
def __repr__(self):
return "Point({0.x!r}, {0.y!r})".format(self)
78 ➤
The built-in repr() function calls the __repr__() special method for the object
it is given and returns the result. The string returned is one of two kinds.
One kind is where the string returned can be evaluated using the built-in
eval() function to produce an object equivalent to the one repr() was called
on. The other kind is used where this is not possible; we will see an example
later on. Here is how we can go from a Point object to a string and back to a
Point object:
Custom Classes
p = Shape.Point(3, 9)
q = eval(p.__module__ + "." + repr(p))
195 ➤
# returns: 'Point(3, 9)'
# returns: 'Point(3, 9)'
We must give the module name when eval()-ing if we used import Shape. (This
would not be necessary if we had done the import differently, for example, from
Shape import Point.) Python provides every object with a few private attributes,
one of which is __module__, a string that holds the object’s module name, which
in this example is "Shape".
At the end of this snippet we have two Point objects, p and q, both with the
same attribute values, so they compare as equal. The eval() function returns
the result of executing the string it is given—which must contain a valid
Python statement.
def __str__(self):
return "({0.x!r}, {0.y!r})".format(self)
The built-in str() function works like the repr() function, except that it calls
the object’s __str__() special method. The result is intended to be understandable to human readers and is not expected to be suitable for passing to the
eval() function. Continuing the previous example, str(p) (or str(q)) would return the string '(3, 9)'.
We have now covered the simple Point class—and also covered a lot of behindthe-scenes details that are important to know but which can mostly be left in
the background. The Point class holds an (x, y) coordinate—a fundamental part
of what we need to represent a circle, as we discussed at the beginning of the
chapter. In the next subsection we will see how to create a custom Circle class,
inheriting from Point so that we don’t have to duplicate the code for the x and
y attributes or for the distance_from_origin() method.
Inheritance and Polymorphism
The Circle class builds on the Point class using inheritance. The Circle class
adds one additional data attribute (radius), and three new methods. It also
reimplements a few of Point’s methods. Here is the complete class definition:
class Circle(Point):
def __init__(self, radius, x=0, y=0):
super().__init__(x, y)
self.radius = radius
def edge_distance_from_origin(self):
return abs(self.distance_from_origin() - self.radius)
Dynamic code
➤ 344
Custom Classes
of its computation. Since the Circle class does not provide an implementation of the distance_from_origin() method, the one provided by the Point base
class will be found and used. Contrast this with the reimplementation of the
__eq__() method. This method compares this circle’s radius with the other circle’s radius, and if they are equal it then explicitly calls the base class’s __eq__()
method using super(). If we did not use super() we would have infinite recursion, since Circle.__eq__() would then just keep calling itself. Notice also that
we don’t have to pass the self argument in the super() calls since Python automatically passes it for us.
Here are a couple of usage examples:
p = Shape.Point(28, 45)
c = Shape.Circle(5, 28, 45)
# returns: 53.0
# returns: 53.0
We can call the distance_from_origin() method on a Point or on a Circle, since
Circles can stand in for Points.
Polymorphism means that any object of a given class can be used as though
it were an object of any of its class’s base classes. This is why when we create
a subclass we need to implement only the additional methods we require and
have to reimplement only those existing methods we want to replace. And
when reimplementing methods, we can use the base class’s implementation if
necessary by using super() inside the reimplementation.
In the Circle’s case we have implemented additional methods, such as area()
and circumference(), and reimplemented methods we needed to change. The
reimplementations of __repr__() and __str__() are necessary because without
them the base class methods will be used and the strings returned will be of
Points instead of Circles. The reimplementations of __init__() and __eq__()
are necessary because we must account for the fact that Circles have an additional attribute, and in both cases we make use of the base class implementations of the methods to minimize the work we must do.
146 ➤
The Point and Circle classes are as complete as we need them to be. We could
provide additional methods, such as other comparison special methods if we
wanted to be able to order Points or Circles. Another thing that we might
want to do for which no method is provided is to copy a Point or Circle. Most
Python classes don’t provide a copy() method (exceptions being dict.copy()
and set.copy()). If we want to copy a Point or Circle we can easily do so by
importing the copy module and using the copy.copy() function. (There is no
need to use copy.deepcopy() for Point and Circle objects since they contain only
immutable instance variables.)
Chapter 6. Object-Oriented Programming
Using Properties to Control Attribute Access
In the previous subsection the Point class included a distance_from_origin()
method, and the Circle class had the area(), circumference(), and edge_distance_from_origin() methods. All these methods return a single float value, so
from the point of view of a user of these classes they could just as well be data
attributes, but read-only, of course. In the file alternative implementations of Point and Circle are provided, and all the methods mentioned
here are provided as properties. This allows us to write code like this:
circle = Shape.Circle(5, 28, 45)
# assumes: import ShapeAlt as Shape
# returns: 5
# returns: 48.0
Here are the implementations of the getter methods for the ShapeAlt.Circle
class’s area and edge_ distance_from_origin properties:
def area(self):
return math.pi * (self.radius ** 2)
def edge_distance_from_origin(self):
return abs(self.distance_from_origin - self.radius)
If we provide only getters as we have done here, the properties are read-only.
The code for the area property is the same as for the previous area() method.
The edge_distance_from_origin’s code is slightly different from before because it
now accesses the base class’s distance_from_origin property instead of calling
a distance_from_origin() method. The most notable difference to both is the
property decorator. A decorator is a function that takes a function or method
as its argument and returns a “decorated” version, that is, a version of the
function or method that is modified in some way. A decorator is indicated by
preceding its name with an at symbol (@). For now, just treat decorators as
syntax—in Chapter 8 we will see how to create custom decorators.
The property() decorator function is built-in and takes up to four arguments: a
getter function, a setter function, a deleter function, and a docstring. The
effect of using @property is the same as calling the property() function with just
one argument, the getter function. We could have created the area property
like this:
def area(self):
return math.pi * (self.radius ** 2)
area = property(area)
We rarely use this syntax, since using a decorator is shorter and clearer.
Custom Classes
In the previous subsection we noted that no validation is performed on the
Circle’s radius attribute. We can provide validation by making radius into a
property. This does not require any changes to the Circle.__init__() method,
and any code that accesses the Circle.radius attribute will continue to work
unchanged—only now the radius will be validated whenever it is set.
Python programmers normally use properties rather than the explicit getters
and setters (e.g., getRadius() and setRadius()) that are so commonly used in
other object-oriented languages. This is because it is so easy to change a data
attribute into a property without affecting the use of the class.
To turn an attribute into a readable/writable property we must create a private
attribute where the data is actually held and supply getter and setter methods.
Here is the radius’s getter, setter, and docstring in full:
def radius(self):
"""The circle's radius
>>> circle = Circle(-2)
Traceback (most recent call
AssertionError: radius must
>>> circle = Circle(4)
>>> circle.radius = -1
Traceback (most recent call
AssertionError: radius must
>>> circle.radius = 6
return self.__radius
be nonzero and non-negative
be nonzero and non-negative
def radius(self, radius):
assert radius > 0, "radius must be nonzero and non-negative"
self.__radius = radius
We use an assert to ensure a nonzero and non-negative radius and store the
radius’s value in the private attribute self.__radius. Notice that the getter and
setter (and deleter if we needed one) all have the same name—it is the decorators that distinguish them, and the decorators rename them appropriately so
that no name conflicts occur.
The decorator for the setter may look strange at first sight. Every property
that is created has a getter, setter, and deleter attribute, so once the radius
property is created using @property, the radius.getter, radius.setter, and
radius.deleter attributes become available. The radius.getter is set to the
Chapter 6. Object-Oriented Programming
getter method by the @property decorator. The other two are set up by Python
so that they do nothing (so the attribute cannot be written to or deleted), unless
they are used as decorators, in which case they in effect replace themselves
with the method they are used to decorate.
The Circle’s initializer, Circle.__init__(), includes the statement self.radius =
radius; this will call the radius property’s setter, so if an invalid radius is given
when a Circle is created an AssertionError exception will be raised. Similarly,
if an attempt is made to set an existing Circle’s radius to an invalid value,
again the setter will be called and an exception raised. The docstring includes
doctests to test that the exception is correctly raised in these cases. (Testing is
covered more fully in Chapter 9.)
Both the Point and Circle types are custom data types that have sufficient
functionality to be useful. Most of the data types that we are likely to create
are like this, but occasionally it is necessary to create a custom data type that
is complete in every respect. We will see examples of this in the next subsection.
Creating Complete Fully Integrated Data Types
When creating a complete data type two possibilities are open to us. One is to
create the data type from scratch. Although the data type will inherit object
(as all Python classes do), every data attribute and method that the data type
requires (apart from __new__()) must be provided. The other possibility is to
inherit from an existing data type that is similar to the one we want to create.
In this case the work usually involves reimplementing those methods we want
to behave differently and “unimplementing” those methods we don’t want
at all.
In the following subsubsection we will implement a FuzzyBool data type from
scratch, and in the subsubsection after that we will implement the same type
but will use inheritance to reduce the work we must do. The built-in bool type
is two-valued (True and False), but in some areas of AI (Artificial Intelligence),
fuzzy Booleans are used, which have values corresponding to “true” and “false”,
and also to intermediates between them. In our implementations we will use
floating-point values with 0.0 denoting False and 1.0 denoting True. In this
system, 0.5 means 50 percent true, and 0.25 means 25 percent true, and so on.
Here are some usage examples (they work the same with either implementation):
a = FuzzyBool.FuzzyBool(.875)
b = FuzzyBool.FuzzyBool(.25)
a >= b
bool(a), bool(b)
# returns: True
# returns: (True, False)
# returns: FuzzyBool(0.125)
Custom Classes
a & b
b |= FuzzyBool.FuzzyBool(.5)
"a={0:.1%} b={1:.0%}".format(a, b)
# returns: FuzzyBool(0.25)
# b is now: FuzzyBool(0.5)
# returns: 'a=87.5% b=50%'
We want the FuzzyBool type to support the complete set of comparison operators (<, <=, ==, !=, >=, >), and the three basic logical operations, not (~), and (&),
and or (|). In addition to the logical operations we want to provide a couple of
other logical methods, conjunction() and disjunction(), that take as many
FuzzyBools as we like and return the appropriate resultant FuzzyBool. And to
complete the data type we want to provide conversions to types bool, int, float,
and str, and have an eval()-able representational form. The final requirements
are that FuzzyBool supports str.format() format specifications, that FuzzyBools
can be used as dictionary keys or as members of sets, and that FuzzyBools are
immutable—but with the provision of augmented assignment operators (&=
and |=) to ensure that they are convenient to use.
Table 6.1 (242 ➤) lists the comparison special methods, Table 6.2 (➤ 250) lists
the fundamental special methods, and Table 6.3 (➤ 253) lists the numeric special methods—these include the bitwise operators (~, &, and |) which FuzzyBools
use for their logical operators, and also arithmetic operators such as + and which FuzzyBool does not implement because they are inappropriate.
Creating Data Types from Scratch
To create the FuzzyBool type from scratch means that we must provide an
attribute to hold the FuzzyBool’s value and all the methods that we require.
Here are the class line and the initializer, taken from
class FuzzyBool:
def __init__(self, value=0.0):
self.__value = value if 0.0 <= value <= 1.0 else 0.0
247 ➤
We have made the value attribute private because we want FuzzyBool to behave
like immutables, so allowing access to the attribute would be wrong. Also, if an
out-of-range value is given we force it to take a fail-safe value of 0.0 (false). In
the previous subsection’s ShapeAlt.Circle class we used a stricter policy, raising
an exception if an invalid radius value was used when creating a new Circle
object. The FuzzyBool’s inheritance tree is shown in Figure 6.4.
The simplest logical operator is logical
bitwise inversion operator (~):
for which we have coopted the
def __invert__(self):
return FuzzyBool(1.0 - self.__value)
Custom Classes
conjunction() # static
Figure 6.4 The FuzzyBool class’s inheritance hierarchy
def __iand__(self, other):
self.__value = min(self.__value, other.__value)
return self
The bitwise AND operator returns a new FuzzyBool based on this one and the
other one, whereas the augmented assignment (in-place) version updates the
private value. Strictly speaking, this is not immutable behavior, but it does
match the behavior of some other Python immutables, such as int, where, for
example, using += looks like the left-hand operand is being changed but in fact
it is re-bound to refer to a new int object that holds the result of the addition.
In this case no rebinding is needed because we really do change the FuzzyBool
itself. And we return self to support the chaining of operations.
We could also implement __rand__(). This method is called when self and other
are of different types and the __and__() method is not implemented for that
particular pair of types. This isn’t needed for the FuzzyBool class. Most of the
special methods for binary operators have both “i” (in-place) and “r” (reflect,
that is, swap operands) versions.
We have not shown the implementation for __or__() which provides the bitwise
| operator, or for __ior__() which provides the in-place |= operator, since both
are the same as the equivalent AND methods except that we take the maximum
value instead of the minimum value of self and other.
def __repr__(self):
return ("{0}({1})".format(self.__class__.__name__,
Chapter 6. Object-Oriented Programming
We have created an eval()-able representational form. For example, given f =
FuzzyBool.FuzzyBool(.75); repr(f) will produce the string 'FuzzyBool(0.75)'.
All objects have some special attributes automatically supplied by Python,
one of which is called __class__, a reference to the object’s class. All classes
have a private __name__ attribute, again provided automatically. We have used
these attributes to provide the class name used for the representation form.
This means that if the FuzzyBool class is subclassed just to add extra methods,
the inherited __repr__() method will work correctly without needing to be
reimplemented, since it will pick up the subclass’s class name.
def __str__(self):
return str(self.__value)
For the string form we just return the floating-point value formatted as a
string. We don’t have to use super() to avoid infinite recursion because we call
str() on the self.__value attribute, not on the instance itself.
def __bool__(self):
return self.__value > 0.5
def __int__(self):
return round(self.__value)
def __float__(self):
return self.__value
The __bool__() special method converts the instance to a Boolean, so it must always return either True or False. The __int__() special method provides integer
conversion. We have used the built-in round() function because int() simply
truncates (so would return 0 for any FuzzyBool value except 1.0). Floating-point
conversion is easy because the value is already a floating-point number.
def __lt__(self, other):
return self.__value < other.__value
def __eq__(self, other):
return self.__value == other.__value
To provide the complete set of comparisons (<, <=, ==, !=, >=, >) it is necessary to
implement at least three of them, <, <=, and ==, since Python can infer > from
<, != from ==, and >= from <=. We have shown only two representative methods
here since all of them are very similar.★
def __hash__(self):
return hash(id(self))
In fact, we implemented only the __lt__() and __eq__() methods quoted here—the other
comparison methods were automatically generated; we will see how in Chapter 8.
➤ 379
Custom Classes
Table 6.3 Numeric and Bitwise Special Methods
Special Method
Special Method
bin(x) oct(x) __round__(self,
__add__(self, other)
x + y
__sub__(self, other)
x - y
__iadd__(self, other)
x += y
__isub__(self, other)
x -= y
__radd__(self, other)
y + x
__rsub__(self, other)
y - x
__mul__(self, other)
x * y
__mod__(self, other)
x % y
__imul__(self, other)
x *= y
__imod__(self, other)
x %= y
__rmul__(self, other)
y * x
__rmod__(self, other)
y % x
x // y
x / y
x //= y
x /= y
y // x
y / x
divmod(x, y)
divmod(y, x)
__pow__(self, other)
x ** y
__and__(self, other)
x & y
__ipow__(self, other)
x **= y
__iand__(self, other)
x &= y
__rpow__(self, other)
y ** x
__rand__(self, other)
y & x
__xor__(self, other)
x ^ y
__or__(self, other)
x | y
__ixor__(self, other)
x ^= y
__ior__(self, other)
x |= y
__rxor__(self, other)
y ^ x
__ror__(self, other)
y | x
x << y
x >> y
x <<= y
x >>= y
y << x
y >> x
Chapter 6. Object-Oriented Programming
By default, instances of custom classes support operator == (which always returns False), and are hashable (so can be dictionary keys and can be added
to sets). But if we reimplement the __eq__() special method to provide proper
equality testing, instances are no longer hashable. This can be fixed by providing a __hash__() special method as we have done here.
Python provides hash functions for strings, numbers, frozen sets, and other
classes. Here we have simply used the built-in hash() function (which can
operate on any type which has a __hash__() special method), and given it the
object’s unique ID from which to calculate the hash. (We can’t use the private
self.__value since that can change as a result of augmented assignment,
whereas an object’s hash value must never change.)
The built-in id() function returns a unique integer for the object it is given
as its argument. This integer is usually the object’s address in memory, but
all that we can assume is that no two objects have the same ID. Behind the
scenes the is operator uses the id() function to determine whether two object
references refer to the same object.
def __format__(self, format_spec):
return format(self.__value, format_spec)
The built-in format() function is only really needed in class definitions. It takes
a single object and an optional format specification and returns a string with
the object suitably formatted.
248 ➤
When an object is used in a format string the object’s __format__() method is
called with the object and the format specification as arguments. The method
returns the instance suitably formatted as we saw earlier.
All the built-in classes already have suitable __format__() methods; here we
make use of the float.__format__() method by passing the floating-point value
and the format string we have been given. We could have achieved exactly the
same thing like this:
def __format__(self, format_spec):
return self.__value.__format__(format_spec)
Using the format() function requires a tiny bit less typing and is clearer to read.
Nothing forces us to use the format() function at all, so we could invent our own
format specification language and interpret it inside the __format__() method,
as long as we return a string.
def conjunction(*fuzzies):
return FuzzyBool(min([float(x) for x in fuzzies]))
Custom Classes
The built-in staticmethod() function is designed to be used as a decorator as we
have done here. Static methods are simply methods that do not get self or any
other first argument specially passed by Python.
The & operator can be chained, so given FuzzyBool’s f, g, and h, we can get the
conjunction of all of them by writing f & g & h. This works fine for small numbers of FuzzyBools, but if we have a dozen or more it starts to become rather
inefficient since each & represents a function call. With the method given
here we can achieve the same thing using a single function call of FuzzyBool.FuzzyBool.conjunction(f, g, h). This can be written more concisely using a FuzzyBool instance, but since static methods don’t get self, if we call
one using an instance and we want to process that instance we must pass it
ourselves—for example, f.conjunction(f, g, h).
We have not shown the corresponding disjunction() method since it differs
only in its name and that it uses max() rather than min().
Some Python programmers consider the use of static methods to be un-Pythonic, and use them only if they are converting code from another language (such
as C++ or Java), or if they have a method that does not use self. In Python,
rather than using static methods it is usually better to create a module function
instead, as we will see in the next subsubsection, or a class method, as we will
see in the last section.
In a similar vein, creating a variable inside a class definition but outside
any method creates a static (class) variable. For constants it is usually more
convenient to use private module globals, but class variables can often be
useful for sharing data among all of a class’s instances.
We have now completed the implementation of the FuzzyBool class “from
scratch”. We have had to reimplement 15 methods (17 if we had done the
minimum of all four comparison operators), and have implemented two static
methods. In the following subsubsection we will show an alternative implementation, this time based on the inheritance of float. It involves the reimplementations of just eight methods and the implementation of two module
functions—and the “unimplementation” of 32 methods.
In most object-oriented languages inheritance is used to create new classes
that have all the methods and attributes of the classes they inherit, as well
as the additional methods and attributes that we want the new class to have.
Python fully supports this, allowing us to add new methods, or to reimplement
inherited methods so as to modify their behavior. But in addition, Python
allows us to effectively unimplement methods, that is, to make the new class
behave as though it does not have some of the methods that it inherits. Doing
this might not appeal to object-oriented purists since it breaks polymorphism,
but in Python at least, it can occasionally be a useful technique.
Chapter 6. Object-Oriented Programming
Creating Data Types from Other Data Types
The FuzzyBool implementation in this subsubsection is in the file One immediate difference from the previous version is that instead
of providing static methods for conjunction() and disjunction(), we have provided them as module functions. For example:
def conjunction(*fuzzies):
return FuzzyBool(min(fuzzies))
The code for this is much simpler than before because FuzzyBoolAlt.FuzzyBool
objects are float subclasses, and so can be used directly in place of a float
without needing any conversion. (The inheritance tree is shown in Figure 6.5.)
Accessing the function is also cleaner than before. Instead of having to specify
both the module and the class (or using an instance), having done import
FuzzyBoolAlt we can just write FuzzyBoolAlt.conjunction().
Figure 6.5 The alternative FuzzyBool class’s inheritance hierarchy
Here is the FuzzyBool’s class line and its __new__() method:
class FuzzyBool(float):
def __new__(cls, value=0.0):
return super().__new__(cls,
value if 0.0 <= value <= 1.0 else 0.0)
When we create a new class it is usually mutable and relies on object.__new__()
to create the raw uninitialized object. But in the case of immutable classes we
need to do the creation and initialization in one step since once an immutable
object has been created it cannot be changed.
Custom Classes
The __new__() method is called before any object has been created (since object
creation is what __new__() does), so it cannot have a self object passed to it
since one doesn’t yet exist. In fact, __new__() is a class method—these are
similar to normal methods except that they are called on the class rather than
on an instance and Python supplies as their first argument the class they are
called on. The variable name cls for class is just a convention, in the same way
that self is the conventional name for the object itself.
So when we write f = FuzzyBool(0.7), under the hood Python calls FuzzyBool.__new__(FuzzyBool, 0.7) to create a new object—say, fuzzy—and then
calls fuzzy.__init__() to do any further initialization, and finally returns an
object reference to the fuzzy object—it is this object reference that f is set to.
Most of __new__()’s work is passed on to the base class implementation, object.__new__(); all we do is make sure that the value is in range.
Class methods are set up by using the built-in classmethod() function used as
a decorator. But as a convenience we don’t have to bother writing @classmethod
before def __new__() because Python already knows that this method is always
a class method. We do need to use the decorator if we want to create other class
methods, though, as we will see in the chapter’s final section.
Now that we have seen a class method we can clarify the different kinds of
methods that Python provides. Class methods have their first argument
added by Python and it is the method’s class; normal methods have their first
argument added by Python and it is the instance the method was called on; and
static methods have no first argument added. And all the kinds of methods get
any arguments we pass to them (as their second and subsequent arguments
in the case of class and normal methods, and as their first and subsequent
arguments for static methods).
def __invert__(self):
return FuzzyBool(1.0 - float(self))
This method is used to provide support for the bitwise NOT operator (~) just
the same as before. Notice that instead of accessing a private attribute that
holds the FuzzyBool’s value we use self directly. This is thanks to inheriting float which means that a FuzzyBool can be used wherever a float is
expected—providing none of the FuzzyBool’s “unimplemented” methods are
used, of course.
def __and__(self, other):
return FuzzyBool(min(self, other))
def __iand__(self, other):
return FuzzyBool(min(self, other))
The logic for these is also the same as before (although the code is subtly
different), and just like the __invert__() method we can use both self and other
Chapter 6. Object-Oriented Programming
directly as though they were floats. We have omitted the OR versions since
they differ only in their names (__or__() and __ior__()) and that they use max()
rather than min().
def __repr__(self):
return ("{0}({1})".format(self.__class__.__name__,
We must reimplement the __repr__() method since the base class version
float.__repr__() just returns the number as a string, whereas we need the class
name to make the representation eval()-able. For the str.format()’s second argument we cannot just pass self since that will result in an infinite recursion
of calls to this __repr__() method, so instead we call the base class implementation.
We don’t have to reimplement the __str__() method because the base class
version, float.__str__(), is sufficient and will be used in the absence of a
FuzzyBool.__str__() reimplementation.
def __bool__(self):
return self > 0.5
def __int__(self):
return round(self)
When a float is used in a Boolean context it is False if its value is 0.0 and True
otherwise. This is not the appropriate behavior for FuzzyBools, so we have had
to reimplement this method. Similarly, using int(self) would simply truncate,
turning everything but 1.0 into 0, so here we use round() to produce 0 for values
up to 0.5 and 1 for values up to and including the maximum of 1.0.
We have not reimplemented the __hash__() method, the __format__() method,
or any of the methods that are used to provide the comparison operators, since
all those provided by the float base class work correctly for FuzzyBools.
The methods we have reimplemented provide a complete implementation of
the FuzzyBool class—and have required far less code than the implementation
presented in the previous subsubsection. However, this new FuzzyBool class
has inherited more than 30 methods which don’t make sense for FuzzyBools.
For example, none of the basic numeric and bitwise shift operators (+, -, *, /, <<,
>>, etc.) can sensibly be applied to FuzzyBools. Here is how we could begin to
“unimplement” addition:
def __add__(self, other):
raise NotImplementedError()
We would also have to write the same code for the __iadd__() and __radd__()
methods to completely prevent addition. (Note that NotImplementedError is a
standard exception and is different from the built-in NotImplemented object.) An
Custom Classes
alternative to raising a NotImplementedError exception, especially if we want
to more closely mimic the behavior of Python’s built-in classes, is to raise
a TypeError. Here is how we can make FuzzyBool.__add__() behave just like
built-in classes that are faced with an invalid operation:
def __add__(self, other):
raise TypeError("unsupported operand type(s) for +: "
"'{0}' and '{1}'".format(
self.__class__.__name__, other.__class__.__name__))
For unary operations, we want to unimplement in a way that mimics the
behavior of built-in types, the code is slightly easier:
def __neg__(self):
raise TypeError("bad operand type for unary -: '{0}'".format(
For comparison operators, there is a much simpler idiom. For example, to
unimplement ==, we would write:
def __eq__(self, other):
return NotImplemented
If a method implementing a comparison operator (<, <=, ==, !=, >=, >), returns
the built-in NotImplemented object and an attempt is made to use the method,
Python will first try the reverse comparison by swapping the operands (in
case the other object has a suitable comparison method since the self object
does not), and if that doesn’t work Python raises a TypeError exception with a
message that explains that the operation is not supported for operands of the
types used. But for all noncomparison methods that we don’t want, we must
raise either a NotImplementedError or a TypeError exception as we did for the
__add__() and __neg__() methods shown earlier.
It would be tedious to unimplement every method we don’t want as we have
done here, although it does work and has the virtue of being easy to understand. Here we will look at a more advanced technique for unimplementing
methods—it is used in the FuzzyBoolAlt module—but it is probably best to skip
to the next section (➤ 261) and return here only if the need arises in practice.
Here is the code for unimplementing the two unary operations we don’t want:
for name, operator in (("__neg__", "-"),
("__index__", "index()")):
message = ("bad operand type for unary {0}: '{{self}}'"
exec("def {0}(self): raise TypeError(\"{1}\".format("
"self=self.__class__.__name__))".format(name, message))
Chapter 6. Object-Oriented Programming
The built-in exec() function dynamically executes the code passed to it from the
object it is given. In this case we have given it a string, but it is also possible to
pass some other kinds of objects. By default, the code is executed in the context
of the enclosing scope, in this case within the definition of the FuzzyBool class,
so the def statements that are executed create FuzzyBool methods which is
what we want. The code is executed just once, when the FuzzyBoolAlt module
is imported. Here is the code that is generated for the first tuple ("__neg__",
def __neg__(self):
raise TypeError("bad operand type for unary -: '{self}'"
We have made the exception and error message match those that Python uses
for its own types. The code for handling binary methods and n-ary functions
(such as pow()) follows a similar pattern but with a different error message. For
completeness, here is the code we have used:
for name, operator in (("__xor__", "^"), ("__ixor__", "^="),
("__add__", "+"), ("__iadd__", "+="), ("__radd__", "+"),
("__sub__", "-"), ("__isub__", "-="), ("__rsub__", "-"),
("__mul__", "*"), ("__imul__", "*="), ("__rmul__", "*"),
("__pow__", "**"), ("__ipow__", "**="),
("__rpow__", "**"), ("__floordiv__", "//"),
("__ifloordiv__", "//="), ("__rfloordiv__", "//"),
("__truediv__", "/"), ("__itruediv__", "/="),
("__rtruediv__", "/"), ("__divmod__", "divmod()"),
("__rdivmod__", "divmod()"), ("__mod__", "%"),
("__imod__", "%="), ("__rmod__", "%"),
("__lshift__", "<<"), ("__ilshift__", "<<="),
("__rlshift__", "<<"), ("__rshift__", ">>"),
("__irshift__", ">>="), ("__rrshift__", ">>")):
message = ("unsupported operand type(s) for {0}: "
"'{{self}}'{{join}} {{args}}".format(operator))
exec("def {0}(self, *args):\n"
types = [\"'\" + arg.__class__.__name__ + \"'\" "
"for arg in args]\n"
raise TypeError(\"{1}\".format("
"self=self.__class__.__name__, "
"join=(\" and\" if len(args) == 1 else \",\"),"
"args=\", \".join(types)))".format(name, message))
This code is slightly more complicated than before because for binary operators
we must output messages where the two types are listed as type1 and type2,
but for three or more types we must list them as type1, type2, type3 to mimic
Dynamic programming
➤ 349
Custom Classes
the built-in behavior. Here is the code that is generated for the first tuple
("__xor__", "^"):
def __xor__(self, *args):
types = ["'" + arg.__class__.__name__ + "'" for arg in args]
raise TypeError("unsupported operand type(s) for ^: "
"'{self}'{join} {args}".format(
join=(" and" if len(args) == 1 else ","),
args=", ".join(types)))
The two for … in loop blocks we have used here can be simply cut and pasted,
and then we can add or remove unary operators and methods from the first
one and binary or n-ary operators and methods from the second one to unimplement whatever methods are not required.
With this last piece of code in place, if we had two FuzzyBools, f and g, and tried
to add them using f + g, we would get a TypeError exception with the message
“unsupported operand type(s) for +: 'FuzzyBool' and 'FuzzyBool'”, which is
exactly the behavior we want.
Creating classes the way we did for the first FuzzyBool implementation is
much more common and is sufficient for almost every purpose. However, if
we need to create an immutable class, the way to do it is to reimplement object.__new__() having inherited one of Python’s immutable types such as
float, int, str, or tuple, and then implement all the other methods we need.
The disadvantage of doing this is that we may need to unimplement some
methods—this breaks polymorphism, so in most cases using aggregation as we
did in the first FuzzyBool implementation is a much better approach.
Custom Collection Classes
In this section’s subsections we will look at custom classes that are responsible
for large amounts of data. The first class we will review, Image, is one that holds
image data. This class is typical of many data-holding custom classes in that it
not only provides in-memory access to its data, but also has methods for saving
and loading the data to and from disk. The second and third classes we will
study, SortedList and SortedDict, are designed to fill a rare and surprising gap
in Python’s standard library for intrinsically sorted collection data types.
Creating Classes That Aggregate Collections
A simple way of representing a 2D color image is as a two-dimensional array
with each array element being a color. So to represent a 100 × 100 image we
must store 10 000 colors. For the Image class (in file, we will take a
Chapter 6. Object-Oriented Programming
potentially more efficient approach. An Image stores a single background color,
plus the colors of those points in the image that differ from the background
color. This is done by using a dictionary as a kind of sparse array, with each key
being an (x, y) coordinate and the corresponding value being the color of that
point. If we had a 100 × 100 image and half its points are the background color,
we would need to store only 5 000 + 1 colors, a considerable saving in memory.
The module follows what should now be a familiar pattern: It starts
with a shebang line, then copyright information in comments, then a module
docstring with some doctests, and then the imports, in this case of the os and
pickle modules. We will briefly cover the use of the pickle module when we
cover saving and loading images. After the imports we create some custom
exception classes:
class ImageError(Exception): pass
class CoordinateError(ImageError): pass
We have shown only the first two exception classes; the others (LoadError,
SaveError, ExportError, and NoFilenameError) are all created the same way and
all inherit from ImageError. Users of the Image class can choose to test for any
of the specific exceptions, or just for the base class ImageError exception.
The rest of the module consists of the Image class and at the end the standard
three lines for running the module’s doctests. Before looking at the class and
its methods, let’s look at how it can be used:
border_color = "#FF0000"
# red
square_color = "#0000FF"
# blue
width, height = 240, 60
midx, midy = width // 2, height // 2
image = Image.Image(width, height, "square_eye.img")
for x in range(width):
for y in range(height):
if x < 5 or x >= width - 5 or y < 5 or y >= height - 5:
image[x, y] = border_color
elif midx - 20 < x < midx + 20 and midy - 20 < y < midy + 20:
image[x, y] = square_color
Notice that we can use the item access operator ([]) for setting colors in the
image. Brackets can also be used for getting or deleting (effectively setting to
the background color) the color at a particular (x, y) coordinate. The coordinates
are passed as a single tuple object (thanks to the comma operator), the same as
if we wrote image[(x, y)]. Achieving this kind of seamless syntax integration
is easy in Python—we just have to implement the appropriate special methods,
➤ 292
Custom Collection Classes
which in the case of the item access operator are __getitem__(), __setitem__(),
and __delitem__().
The Image class uses HTML-style hexadecimal strings to represent colors. The
background color must be set when the image is created; otherwise, it defaults
to white. The Image class saves and loads images in its own custom format,
but it can also export in the .xpm format which is understood by many image
processing applications. The .xpm image produced by the code snippet is shown
in Figure 6.6.
Figure 6.6 The square_eye.xpm image
We will now review the Image class’s methods, starting with the class line and
the initializer:
class Image:
def __init__(self, width, height, filename="",
self.filename = filename
self.__background = background
self.__data = {}
self.__width = width
self.__height = height
self.__colors = {self.__background}
When an Image is created, the user (i.e., the class’s user) must provide a width
and height, but the filename and background color are optional since defaults
are provided. The self.__data dictionary’s keys are (x, y) coordinates and its values are color strings. The self.__colors set is initialized with the background
color; it is used to keep track of the unique colors used by the image.
All the data attributes are private except for the filename, so we must provide
a means by which users of the class can access them. This is easily done using
def background(self):
return self.__background
In Chapter 8 we will see a completely different approach to providing attribute access, using
special methods such as __getattr__() and __setattr__(), that is useful in some circumstances.
Chapter 6. Object-Oriented Programming
def width(self):
return self.__width
def height(self):
return self.__height
def colors(self):
return set(self.__colors)
146 ➤
When returning a data attribute from an object we need to be aware of whether
the attribute is of an immutable or mutable type. It is always safe to return immutable attributes since they can’t be changed, but for mutable attributes we
must consider some trade-offs. Returning a reference to a mutable attribute is
very fast and efficient because no copying takes place—but it also means that
the caller now has access to the object’s internal state and might change it in
a way that invalidates the object. One policy to consider is to always return a
copy of mutable data attributes, unless profiling shows a significant negative
effect on performance. (In this case, an alternative to keeping the set of unique
colors would be to return set(self.__data.values()) | {self.__background}
whenever the set of colors was needed; we will return to this theme shortly.)
def __getitem__(self, coordinate):
assert len(coordinate) == 2, "coordinate should be a 2-tuple"
if (not (0 <= coordinate[0] < self.width) or
not (0 <= coordinate[1] < self.height)):
raise CoordinateError(str(coordinate))
return self.__data.get(tuple(coordinate), self.__background)
This method returns the color for a given coordinate using the item access
operator ([]). The special methods for the item access operators and some other
collection-relevant special methods are listed in Table 6.4.
We have chosen to apply two policies for item access. The first policy is that a
precondition for using an item access method is that the coordinate it is passed
is a sequence of length 2 (usually a 2-tuple), and we use an assertion to ensure
this. The second policy is that any coordinate values are accepted, but if either
is out of range, we raise a custom exception.
We have used the dict.get() method with a default value of the background
color to retrieve the color for the given coordinate. This ensures that if the color
has never been set for the coordinate the background color is correctly returned
instead of a KeyError exception being raised.
Custom Collection Classes
Table 6.4 Collection Special Methods
Special Method
__contains__(self, x)
x in y
Returns True if x is in sequence y or if
x is a key in mapping y
__delitem__(self, k)
del y[k]
Deletes the k-th item of sequence y or
the item with key k in mapping y
__getitem__(self, k)
Returns the k-th item of sequence y or
the value for key k in mapping y
for x in y: Returns an iterator for sequence y’s
items or mapping y’s keys
reversed(y) Returns a backward iterator for sequence y’s items or mapping y’s keys
__setitem__(self, k, v) y[k] = v
Returns the number of items in y
Sets the k-th item of sequence y or the
value for key k in mapping y, to v
def __setitem__(self, coordinate, color):
assert len(coordinate) == 2, "coordinate should be a 2-tuple"
if (not (0 <= coordinate[0] < self.width) or
not (0 <= coordinate[1] < self.height)):
raise CoordinateError(str(coordinate))
if color == self.__background:
self.__data.pop(tuple(coordinate), None)
self.__data[tuple(coordinate)] = color
If the user sets a coordinate’s value to the background color we can simply
delete the corresponding dictionary item since any coordinate not in the dictionary is assumed to have the background color. We must use dict.pop() and
give a dummy second argument rather than use del because doing so avoids a
KeyError being raised if the key (coordinate) is not in the dictionary.
If the color is different from the background color, we set it for the given
coordinate and add it to the set of the unique colors used by the image.
def __delitem__(self, coordinate):
assert len(coordinate) == 2, "coordinate should be a 2-tuple"
if (not (0 <= coordinate[0] < self.width) or
not (0 <= coordinate[1] < self.height)):
raise CoordinateError(str(coordinate))
self.__data.pop(tuple(coordinate), None)
Chapter 6. Object-Oriented Programming
If a coordinate’s color is deleted the effect is to make that coordinate’s color
the background color. Again we use dict.pop() to remove the item since it
will work correctly whether or not an item with the given coordinate is in
the dictionary.
Both __setitem__() and __delitem__() have the potential to make the set of
colors contain more colors than the image actually uses. For example, if a
unique nonbackground color is deleted at a certain pixel, the color remains in
the color set even though it is no longer used. Similarly, if a pixel has a unique
nonbackground color and is set to the background color, the unique color is
no longer used, but remains in the color set. This means that, at worst, the
color set could contain more colors than are actually used by the image (but
never less).
We have chosen to accept the trade-off of potentially having more colors in the
color set than are actually used for the sake of better performance, that is, to
make setting and deleting a color as fast as possible—especially since storing
a few more colors isn’t usually a problem. Of course, if we wanted to ensure
that the set of colors was in sync we could either create an additional method
that could be called whenever we wanted, or accept the overhead and do the
computation automatically when it was needed. In either case, the code is very
simple (and is used when a new image is loaded):
self.__colors = (set(self.__data.values()) |
This simply overwrites the set of colors with the set of colors actually used in
the image unioned with the background color.
We have not provided a __len__() implementation since it does not make sense
for a two-dimensional object. Also, we cannot provide a representational form
since an Image cannot be created fully formed just by calling Image(), so we do
not provide __repr__() (or __str__()) implementations either. If a user calls
repr() or str() on an Image object, the object.__repr__() base class implementation will return a suitable string, for example, '<Image.Image object at
0x9c794ac>'. This is a standard format used for non-eval()-able objects. The
hexadecimal number is the object’s ID—this is unique (normally it is the object’s address in memory), but transient.
We want users of the Image class to be able to save and load their image data,
so we have provided two methods, save() and load(), to carry out these tasks.
We have chosen to save the data by pickling it. In Python-speak pickling is
a way of serializing (converting into a sequence of bytes, or into a string) a
Python object. What is so powerful about pickling is that the pickled object
can be a collection data type, such as a list or a dictionary, and even if the
pickled object has other objects inside it (including other collections, which
Chapter 6. Object-Oriented Programming
we had to open the file in binary mode. When reading pickles no protocol is
specified—the pickle.load() function is smart enough to work out the protocol
for itself.
def load(self, filename=None):
if filename is not None:
self.filename = filename
if not self.filename:
raise NoFilenameError()
fh = None
fh = open(self.filename, "rb")
data = pickle.load(fh)
(self.__width, self.__height, self.__background,
self.__data) = data
self.__colors = (set(self.__data.values()) |
except (EnvironmentError, pickle.UnpicklingError) as err:
raise LoadError(str(err))
if fh is not None:
This function starts off the same as the save() function to get the filename of
the file to load. The file must be opened in read binary mode, and the data is
read using the single statement, data = pickle.load(fh). The data object is an
exact reconstruction of the one we saved, so in this case it is a list with the
width and height integers, the background color string, and the dictionary
of coordinate–color items. We use tuple unpacking to assign each of the data
list’s items to the appropriate variable, so any previously held image data is
(correctly) lost.
The set of unique colors is reconstructed by making a set of all the colors in the
coordinate–color dictionary and then adding the background color.
def export(self, filename):
if filename.lower().endswith(".xpm"):
raise ExportError("unsupported export format: " +
We have provided one generic export method that uses the file extension to
determine which private method to call—or raises an exception for file formats
that cannot be exported. In this case we only support saving to .xpm files (and
then only for images with fewer than 8 930 colors). We haven’t quoted the
Custom Collection Classes
__export_xpm() method because it isn’t really relevant to this chapter’s theme,
but it is in the book’s source code, of course.
We have now completed our coverage of the custom Image class. This class is
typical of those used to hold program-specific data, providing access to the
data items it contains, the ability to save and load all its data to and from
disk, and with only the essential methods it needs provided. In the next two
subsections we will see how to create two generic custom collection types that
offer complete APIs.
Creating Collection Classes Using Aggregation
In this subsection we will develop a complete custom collection data type, SortedList, that holds a list of items in sorted order. The items are sorted using
their less than operator (<), provided by the __lt__() special method, or by using a key function if one is given. The class tries to match the API of the builtin list class to make it as easy to learn and use as possible, but some methods
cannot sensibly be provided—for example, using the concatenation operator (+)
could result in items being out of order, so we do not implement it.
As always when creating custom classes, we are faced with the choices of
inheriting a class that is similar to the one we want to make, or creating a class
from scratch and aggregating instances of any other classes we need inside it,
or doing a mixture of both. For this subsection’s SortedList we use aggregation
(and implicitly inherit object, of course), and for the following subsection’s
SortedDict we will use both aggregation and inheritance (inheriting dict).
In Chapter 8 we will see that classes can make promises about the API they
offer. For example, a list provides the MutableSequence API which means that
it supports the in operator, the iter() and len() built-in functions, and the item
access operator ([]) for getting, setting, and deleting items, and an insert()
method. The SortedList class implemented here does not support item setting
and does not have an insert() method, so it does not provide a MutableSequence
API. If we were to create SortedList by inheriting list, the resultant class
would claim to be a mutable sequence but would not have the complete API.
In view of this the SortedList does not inherit list and so makes no promises
about its API. On the other hand, the next subsection’s SortedDict class supports the complete MutableMapping API that the dict class provides, so we can
make it a dict subclass.
Here are some basic examples of using a SortedList:
letters = SortedList.SortedList(("H", "c", "B", "G", "e"), str.lower)
# str(letters) == "['B', 'c', 'e', 'G', 'H']"
Chapter 6. Object-Oriented Programming
# str(letters) == "['A', 'B', 'c', 'e', 'f', 'G', 'G', 'H']"
letters[2] # returns: 'c'
A SortedList object aggregates (is composed of) two private attributes; a function, self.__key() (held as object reference self.__key), and a list, self.__list.
182 ➤
The key function is passed as the second argument (or using the key keyword
argument if no initial sequence is given). If no key function is specified the
following private module function is used:
_identity = lambda x: x
This is the identity function: It simply returns its argument unchanged, so
when it is used as a SortedList’s key function it means that the sort key for each
object in the list is the object itself.
The SortedList type does not allow the item access operator ([]) to change an
item (so it does not implement the __setitem__() special method), nor does
it provide the append() or extend() method since these might invalidate the
ordering. The only way to add items is to pass a sequence when the SortedList
is created or to add them later using the SortedList.add() method. On the other
hand, we can safely use the item access operator for getting or deleting the
item at a given index position since neither operation affects the ordering, so
both the __getitem__() and __delitem__() special methods are implemented.
We will now review the class method by method, starting as usual with the
class line and the initializer:
class SortedList:
def __init__(self, sequence=None, key=None):
self.__key = key or _identity
assert hasattr(self.__key, "__call__")
if sequence is None:
self.__list = []
elif (isinstance(sequence, SortedList) and
sequence.key == self.__key):
self.__list = sequence.__list[:]
self.__list = sorted(list(sequence), key=self.__key)
Since a function’s name is an object reference (to its function), we can hold
functions in variables just like any other object reference. Here the private
self.__key variable holds a reference to the key function that was passed in, or
to the identity function. The method’s first statement relies on the fact that the
or operator returns its first operand if it is True in a Boolean context (which a
not-None key function is), or its second operand otherwise. A slightly longer but
Custom Collection Classes
more obvious alternative would have been self.__key = key if key is not None
else _identity.
Once we have the key function, we use an assert to ensure that it is callable.
The built-in hasattr() function returns True if the object passed as its first argument has the attribute whose name is passed as its second argument. There
are corresponding setattr() and delattr() functions—these functions are covered in Chapter 8. All callable objects, for example, functions and methods,
have a __call__ attribute.
To make the creation of SortedLists as similar as possible to the creation of
lists we have an optional sequence argument that corresponds to the single
optional argument that list() accepts. The SortedList class aggregates a
list collection in the private variable self.__list and keeps the items in the
aggregated list in sorted order using the given key function.
The elif clause uses type testing to see whether the given sequence is a SortedList and if that is the case whether it has the same key function as this sorted list. If these conditions are met we simply shallow-copy the sequence’s list
without needing to sort it. If most key functions are created on the fly using
lambda, even though two may have the same code they will not compare as
equal, so the efficiency gain may not be realized in practice.
def key(self):
return self.__key
Once a sorted list is created its key function is fixed, so we keep it as a private
variable to prevent users from changing it. But some users may want to get a
reference to the key function (as we will see in the next subsection), and so we
have made it accessible by providing the read-only key property.
def add(self, value):
index = self.__bisect_left(value)
if index == len(self.__list):
self.__list.insert(index, value)
When this method is called the given value must be inserted into the private
self.__list in the correct position to preserve the list’s order. The private
SortedList.__bisect_left() method returns the required index position as we
will see in a moment. If the new value is larger than any other value in the list
it must go at the end, so the index position will be equal to the list’s length (list
index positions go from 0 to len(L) - 1)—if this is the case we append the new
value. Otherwise, we insert the new value at the given index position—which
will be at index position 0 if the new value is smaller than any other value in
the list.
Chapter 6. Object-Oriented Programming
def __bisect_left(self, value):
key = self.__key(value)
left, right = 0, len(self.__list)
while left < right:
middle = (left + right) // 2
if self.__key(self.__list[middle]) < key:
left = middle + 1
right = middle
return left
This private method calculates the index position where the given value belongs in the list, that is, the index position where the value is (if it is in the list),
or where it should go (if it isn’t in the list). It computes the comparison key
for the given value using the sorted list’s key function, and compares the comparison key with the computed comparison keys of the items that the method
examines. The algorithm used is binary search (also called binary chop), which
has excellent performance even on very large lists—for example, at most, 21
comparisons are required to find a value’s position in a list of 1 000 000 items.★
Compare this with a plain unsorted list which uses linear search and needs an
average of 500 000 comparisons, and at worst 1 000 000 comparisons, to find a
value in a list of 1 000 000 items.
def remove(self, value):
index = self.__bisect_left(value)
if index < len(self.__list) and self.__list[index] == value:
del self.__list[index]
raise ValueError("{0}.remove(x): x not in list".format(
This method is used to remove the first occurrence of the given value. It uses
the SortedList.__bisect_left() method to find the index position where the
value belongs and then tests to see whether that index position is within the
list and that the item at that position is the same as the given value. If the
conditions are met the item is removed; otherwise, a ValueError exception is
raised (which is what list.remove() does in the same circumstances).
def remove_every(self, value):
count = 0
index = self.__bisect_left(value)
while (index < len(self.__list) and
self.__list[index] == value):
Python’s bisect module provides the bisect.bisect_left() function and some others, but at the
time of this writing none of the bisect module’s functions can work with a key function.
Custom Collection Classes
del self.__list[index]
count += 1
return count
This method is similar to the SortedList.remove() method, and is an extension
of the list API. It starts off by finding the index position where the first
occurrence of the value belongs in the list, and then loops as long as the index
position is within the list and the item at the index position is the same as the
given value. The code is slightly subtle since at each iteration the matching
item is deleted, and as a consequence, after each deletion the item at the index
position is the item that followed the deleted item.
def count(self, value):
count = 0
index = self.__bisect_left(value)
while (index < len(self.__list) and
self.__list[index] == value):
index += 1
count += 1
return count
This method returns the number of times the given value occurs in the list
(which could be 0). It uses a very similar algorithm to SortedList.remove_
every(), only here we must increment the index position in each iteration.
def index(self, value):
index = self.__bisect_left(value)
if index < len(self.__list) and self.__list[index] == value:
return index
raise ValueError("{0}.index(x): x not in list".format(
Since a SortedList is ordered we can use a fast binary search to find (or not find)
the value in the list.
def __delitem__(self, index):
del self.__list[index]
The __delitem__() special method provides support for the del L[n] syntax,
where L is a sorted list and n is an integer index position. We don’t test for an
out-of-range index since if one is given the self.__list[index] call will raise an
IndexError exception, which is the behavior we want.
def __getitem__(self, index):
return self.__list[index]
This method provides support for the x = L[n] syntax, where L is a sorted list
and n is an integer index position.
Chapter 6. Object-Oriented Programming
def __setitem__(self, index, value):
raise TypeError("use add() to insert a value and rely on "
"the list to put it in the right place")
We don’t want the user to change an item at a given index position (so L[n] =
x is disallowed); otherwise, the sorted list’s order might be invalidated. The
TypeError exception is the one used to signify that an operation is not supported
by a particular data type.
def __iter__(self):
return iter(self.__list)
This method is easy to implement since we can just return an iterator to the
private list using the built-in iter() function. This method is used to support
the for value in iterable syntax.
Note that if a sequence is required it is this method that is used. So to convert
a SortedList, L, to a plain list we can call list(L), and behind the scenes
Python will call SortedList.__iter__(L) to provide the sequence that the list()
function requires.
def __reversed__(self):
return reversed(self.__list)
This provides support for the built-in reversed() function so that we can write,
for example, for value in reversed(iterable).
def __contains__(self, value):
index = self.__bisect_left(value)
return (index < len(self.__list) and
self.__list[index] == value)
The __contains__() method provides support for the in operator. Once again we
are able to use a fast binary search rather than the slow linear search used by
a plain list.
def clear(self):
self.__list = []
def pop(self, index=-1):
return self.__list.pop(index)
def __len__(self):
return len(self.__list)
def __str__(self):
return str(self.__list)
Custom Collection Classes
The SortedList.clear() method discards the existing list and replaces it with a
new empty list. The SortedList.pop() method removes and returns the item at
the given index position, or raises an IndexError exception if the index is out of
range. For the pop(), __len__(), and __str__() methods, we simply pass on the
work to the aggregated self.__list object.
We do not reimplement the __repr__() special method, so the base class object.__repr__() will be called when the user writes repr(L) and L is a SortedList. This will produce a string such as '<SortedList.SortedList object at
0x97e7cec>', although the hexadecimal ID will vary, of course. We cannot
provide a sensible __repr__() implementation because we would need to give
the key function and we cannot represent a function object reference as an
eval()-able string.
We have not implemented the insert(), reverse(), or sort() method because
none of them is appropriate. If any of them are called an AttributeError
exception will be raised.
If we copy a sorted list using the L[:] idiom we will get a plain list object,
rather than a SortedList. The easiest way to get a copy is to import the copy
module and use the copy.copy() function—this is smart enough to copy a sorted
list (and instances of most other custom classes) without any help. However,
we have decided to provide an explicit copy() method:
def copy(self):
return SortedList(self, self.__key)
By passing self as the first argument we ensure that self.__list is simply
shallow-copied rather than being copied and re-sorted. (This is thanks to the
__init__() method’s type-testing elif clause.) The theoretical performance
advantage of copying this way is not available to the copy.copy() function, but
we can easily make it available by adding this line:
__copy__ = copy
When copy.copy() is called it tries to use the object’s __copy__() special method,
falling back to its own code if one isn’t provided. With this line in place
copy.copy() will now use the SortedList.copy() method for sorted lists. (It is
also possible to provide a __deepcopy__() special method, but this is slightly
more involved—the copy module’s online documentation has the details.)
We have now completed the implementation of the SortedList class. In the
next subsection we will make use of a SortedList to provide a sorted list of keys
for the SortedDict class.
Chapter 6. Object-Oriented Programming
Creating Collection Classes Using Inheritance
136 ➤
The SortedDict class shown in this subsection attempts to mimic a dict as
closely as possible. The major difference is that a SortedDict’s keys are always
ordered based on a specified key function or on the identity function. SortedDict provides the same API as dict (except for having a non-eval()-able repr()),
plus two extra methods that make sense only for an ordered collection.★ (Note
that Python 3.1 introduced the collections.OrderedDict class—this class is different from SortedDict since it is insertion-ordered rather than key-ordered.)
Here are a few examples of use to give a flavor of how SortedDict works:
d = SortedDict.SortedDict(dict(s=1, A=2, y=6), str.lower)
d["z"] = 4
d["T"] = 5
del d["y"]
d["n"] = 3
d["A"] = 17
str(d) # returns: "{'A': 17, 'n': 3, 's': 1, 'T': 5, 'z': 4}"
The SortedDict implementation uses both aggregation and inheritance. The
sorted list of keys is aggregated as an instance variable, whereas the SortedDict
class itself inherits the dict class. We will start our code review by looking at
the class line and the initializer, and then we will look at all of the other methods in turn.
class SortedDict(dict):
def __init__(self, dictionary=None, key=None, **kwargs):
dictionary = dictionary or {}
if kwargs:
self.__keys = SortedList.SortedList(super().keys(), key)
The dict base class is specified in the class line. The initializer tries to mimic
the dict() function, but adds a second argument for the key function. The
super().__init__() call is used to initialize the SortedDict using the base class
dict.__init__() method. Similarly, if keyword arguments have been used, we
use the base class dict.update() method to add them to the dictionary. (Note
that only one occurrence of any keyword argument is accepted, so none of the
keys in the kwargs keyword arguments can be “dictionary” or “key”.)
The SortedDict class presented here is different from the one in Rapid GUI Programming with
Python and Qt by this author, ISBN 0132354187, and from the one in the Python Package Index.
Custom Collection Classes
We keep a copy of all the dictionary’s keys in a sorted list stored in the
self.__keys variable. We pass the dictionary’s keys to initialize the sorted list
using the base class’s dict.keys() method—we must not use SortedDict.keys()
because that relies on the self.__keys variable which will exist only after the
SortedList of keys has been created.
def update(self, dictionary=None, **kwargs):
if dictionary is None:
elif isinstance(dictionary, dict):
for key, value in dictionary.items():
super().__setitem__(key, value)
if kwargs:
self.__keys = SortedList.SortedList(super().keys(),
This method is used to update one dictionary’s items with another dictionary’s
items, or with keyword arguments, or both. Items which exist only in the
other dictionary are added to this one, and for items whose keys appear in both
dictionaries, the other dictionary’s value replaces the original value. We have
had to extend the behavior slightly in that we keep the original dictionary’s key
function, even if the other dictionary is a SortedDict.
The updating is done in two phases. First we update the dictionary’s items. If
the given dictionary is a dict subclass (which includes SortedDict, of course),
we use the base class dict.update() to perform the update—using the base
class version is essential to avoid calling SortedDict.update() recursively and
going into an infinite loop. If the dictionary is not a dict we iterate over the
dictionary’s items and set each key–value pair individually. (If the dictionary
object is not a dict and does not have an items() method an AttributeError
exception will quite rightly be raised.) If keyword arguments have been used
we again call the base class update() method to incorporate them.
A consequence of the updating is that the self.__keys list becomes out of
date, so we replace it with a new SortedList with the dictionary’s keys (again
obtained from the base class, since the SortedDict.keys() method relies on the
self.__keys list which we are in the process of updating), and with the original
sorted list’s key function.
def fromkeys(cls, iterable, value=None, key=None):
return cls({k: value for k in iterable}, key)
Chapter 6. Object-Oriented Programming
The dict API includes the dict.fromkeys() class method. This method is used
to create a new dictionary based on an iterable. Each element in the iterable
becomes a key, and each key’s value is either None or the specified value.
Because this is a class method the first argument is provided automatically by
Python and is the class. For a dict the class will be dict, and for a SortedDict it
is SortedDict. The return value is a dictionary of the given class. For example:
class MyDict(SortedDict.SortedDict): pass
d = MyDict.fromkeys("VEINS", 3)
# returns: "{'E': 3, 'I': 3, 'N': 3, 'S': 3, 'V': 3}"
# returns: 'MyDict'
So when inherited class methods are called, their cls variable is set to the
correct class, just like when normal methods are called and their self variable
is set to the current object. This is different from and better than using a
static method because a static method is tied to a particular class and does not
know whether it is being executed in the context of its original class or that of
a subclass.
def __setitem__(self, key, value):
if key not in self:
return super().__setitem__(key, value)
This method implements the d[key] = value syntax. If the key isn’t in the
dictionary we add it to the list of keys, relying on the SortedList to put it in the
right place. Then we call the base class method, and return its result to the
caller to support chaining, for example, x = d[key] = value.
Notice that in the if statement we check to see whether the key already exists
in the SortedDict by using not in self. Because SortedDict inherits dict, a
SortedDict can be used wherever a dict is expected, and in this case self is a
SortedDict. When we reimplement dict methods in SortedDict, if we need to
call the base class implementation to get it to do some of the work for us, we
must be careful to call the method using super(), as we do in this method’s last
statement; doing so prevents the reimplementation of the method from calling
itself and going into infinite recursion.
We do not reimplement the __getitem__() method since the base class version
works fine and has no effect on the ordering of the keys.
def __delitem__(self, key):
except ValueError:
raise KeyError(key)
return super().__delitem__(key)
Custom Collection Classes
Generator Functions
A generator function or generator method is one which contains a yield expression. When a generator function is called it returns an iterator. Values
are extracted from the iterator one at a time by calling its __next__() method.
At each call to __next__() the generator function’s yield expression’s value
(None if none is specified) is returned. If the generator function finishes or
executes a return a StopIteration exception is raised.
In practice we rarely call __next__() or catch a StopIteration. Instead, we
just use a generator like any other iterable. Here are two almost equivalent
functions. The one on the left returns a list and the one on the right returns
a generator.
# Build and return a list
def letter_range(a, z):
result = []
while ord(a) < ord(z):
a = chr(ord(a) + 1)
return result
# Return each value on demand
def letter_range(a, z):
while ord(a) < ord(z):
yield a
a = chr(ord(a) + 1)
We can iterate over the result produced by either function using a for loop,
for example, for letter in letter_range("m", "v"):. But if we want a list of
the resultant letters, although calling letter_range("m", "v") is sufficient for
the left-hand function, for the right-hand generator function we must use
list(letter_range("m", "v")).
Generator functions and methods (and generator expressions) are covered
more fully in Chapter 8.
This method provides the del d[key] syntax. If the key is not present the SortedList.remove() call will raise a ValueError exception. If this occurs we catch
the exception and raise a KeyError exception instead so as to match the dict
class’s API. Otherwise, we return the result of calling the base class implementation to delete the item with the given key from the dictionary itself.
def setdefault(self, key, value=None):
if key not in self:
return super().setdefault(key, value)
This method returns the value for the given key if the key is in the dictionary;
otherwise, it creates a new item with the given key and value and returns the
value. For the SortedDict we must make sure that the key is added to the keys
list if the key is not already in the dictionary.
Chapter 6. Object-Oriented Programming
def pop(self, key, *args):
if key not in self:
if len(args) == 0:
raise KeyError(key)
return args[0]
return super().pop(key, args)
If the given key is in the dictionary this method returns the corresponding
value and removes the key–value item from the dictionary. The key must also
be removed from the keys list.
The implementation is quite subtle because the pop() method must support
two different behaviors to match dict.pop(). The first is d.pop(k); here the value
for key k is returned, or if there is no key k, a KeyError is raised. The second is
d.pop(k, value); here the value for key k is returned, or if there is no key k, value
(which could be None) is returned. In all cases, if key k exists, the corresponding
item is removed.
def popitem(self):
item = super().popitem()
return item
The dict.popitem() method removes and returns a random key–value item
from the dictionary. We must call the base class version first since we don’t
know in advance which item will be removed. We remove the item’s key from
the keys list, and then return the item.
def clear(self):
Here we clear all the dictionary’s items and all the keys list’s items.
def values(self):
for key in self.__keys:
yield self[key]
def items(self):
for key in self.__keys:
yield (key, self[key])
def __iter__(self):
return iter(self.__keys)
keys = __iter__
Custom Collection Classes
Dictionaries have four methods that return iterators: dict.values() for the dictionary’s values, dict.items() for the dictionary’s key–value items, dict.keys()
for the keys, and the __iter__() special method that provides support for the
iter(d) syntax, and operates on the keys. (Actually, the base class versions of
these methods return dictionary views, but for most purposes the behavior of
the iterators implemented here is the same.)
Since the __iter__() method and the keys() method have identical behavior,
instead of implementing keys(), we simply create an object reference called
keys and set it to refer to the __iter__() method. With this in place, users of
SortedDict can call d.keys() or iter(d) to get an iterator over a dictionary’s keys,
just the same as they can call d.values() to get an iterator over the dictionary’s
The values() and items() methods are generator methods—see the sidebar
“Generator Functions” (279 ➤) for a brief explanation of generator methods.
In both cases they iterate over the sorted keys list, so they always return iterators that iterate in key order (with the key order depending on the key function given to the initializer). For the items() and values() methods, the values
are looked up using the d[k] syntax (which uses dict.__getitem__() under the
hood), since we can treat self as a dict.
def __repr__(self):
return object.__repr__(self)
def __str__(self):
return ("{" + ", ".join(["{0!r}: {1!r}".format(k, v)
for k, v in self.items()]) + "}")
We cannot provide an eval()-able representation of a SortedDict because we
can’t produce an eval()-able representation of the key function. So for the
__repr__() reimplementation we bypass dict.__repr__(), and instead call
the ultimate base class version, object.__repr__(). This produces a string
of the kind used for non-eval()-able representations, for example, '<SortedDict.SortedDict object at 0xb71fff5c>'.
We have implemented the SortedDict.__str__() method ourselves because we
want the output to show the items in key sorted order. The method could have
been written like this instead:
items = []
for key, value in self.items():
items.append("{0!r}: {1!r}".format(key, value))
return "{" + ", ".join(items) + "}"
Using a list comprehension is shorter and avoids the need for the temporary
items variable.
➤ 341
Chapter 6. Object-Oriented Programming
The base class methods dict.get(), dict.__getitem__() (for the v = d[k] syntax),
dict.__len__() (for len(d)), and dict.__contains__() (for x in d) all work fine as
they are and don’t affect the key ordering, so we have not needed to reimplement them.
The last dict method that we must reimplement is copy().
def copy(self):
d = SortedDict()
super(SortedDict, d).update(self)
d.__keys = self.__keys.copy()
return d
The easiest reimplementation is simply def copy(self): return SortedDict(
self). We’ve chosen a slightly more complicated solution that avoids re-sorting the already sorted keys. We create an empty sorted dictionary, then update it with the items in the original sorted dictionary using the base class
dict.update() to avoid the SortedDict.update() reimplementation, and replace the dictionary’s self.__keys SortedList with a shallow copy of the original one.
When super() is called with no arguments it works on the base class and the
self object. But we can make it work on any class and any object by passing
in a class and an object explicitly. Using this syntax, the super() call works on
the immediate base class of the class it is given, so in this case the code has the
same effect as (and could be written as) dict.update(d, self).
In view of the fact that Python’s sort algorithm is very fast, and is particularly
well optimized for partially sorted lists, the efficiency gain is likely to be little
or nothing except for huge dictionaries. However, the implementation shows
that at least in principle, a custom copy() method can be more efficient than
using the copy_of_x = ClassOfX(x) idiom that Python’s built-in types support.
And just as we did for SortedList, we have set __copy__ = copy so that the
copy.copy() function uses our custom copy method rather than its own code.
def value_at(self, index):
return self[self.__keys[index]]
def set_value_at(self, index, value):
self[self.__keys[index]] = value
These two methods represent an extension to the dict API. Since, unlike a plain
dict, a SortedDict is ordered, it follows that the concept of key index positions
is applicable. For example, the first item in the dictionary is at index position 0,
and the last at position len(d) - 1. Both of these methods operate on the dictionary item whose key is at the index-th position in the sorted keys list. Thanks
to inheritance, we can look up values in the SortedDict using the item access op-
Custom Collection Classes
erator ([]) applied directly to self, since self is a dict. If an out-of-range index
is given the methods raise an IndexError exception.
We have now completed the implementation of the SortedDict class. It is not
often that we need to create complete generic collection classes like this, but
when we do, Python’s special methods allow us to fully integrate our class so
that its users can treat it like any of the built-in or standard library classes.
This chapter covered all the fundamentals of Python’s support for object-oriented programming. We began by showing some of the disadvantages of a purely
procedural approach and how these could be avoided by using object orientation. We then described some of the most common terminology used in objectoriented programming, including many “duplicate” terms such as base class
and super class.
We saw how to create simple classes with data attributes and custom methods.
We also saw how to inherit classes and how to add additional data attributes
and additional methods, and how methods can be “unimplemented”. Unimplementing is needed when we inherit a class but want to restrict the methods
that our subclass provides, but it should be used with care since it breaks the
expectation that a subclass can be used wherever one of its base classes can be
used, that is, it breaks polymorphism.
Custom classes can be seamlessly integrated so that they support the same
syntaxes as Python’s built-in and library classes. This is achieved by implementing special methods. We saw how to implement special methods to support comparisons, how to provide representational and string forms, and how to
provide conversions to other types such as int and float when it makes sense to
do so. We also saw how to implement the __hash__() method to make a custom
class’s instances usable as dictionary keys or as members of a set.
Data attributes by themselves provide no mechanism for ensuring that they
are set to valid values. We saw how easy it is to replace data attributes with
properties—this allows us to create read-only properties, and for writable
properties makes it easy to provide validation.
Most of the classes we create are “incomplete” since we tend to provide only the
methods that we actually need. This works fine in Python, but in addition it is
possible to create complete custom classes that provide every relevant method.
We saw how to do this for single valued classes, both by using aggregation
and more compactly by using inheritance. We also saw how to do this for
multivalued (collection) classes. Custom collection classes can provide the
same facilities as the built-in collection classes, including support for in, len(),
iter(), reversed(), and the item access operator ([]).
Chapter 6. Object-Oriented Programming
We learned that object creation and initialization are separate operations and
that Python allows us to control both, although in almost every case we only
need to customize initialization. We also learned that although it is always
safe to return an object’s immutable data attributes, we should normally only
ever return copies of an object’s mutable data attributes to avoid the object’s
internal state leaking out and being accidentally invalidated.
Python provides normal methods, static methods, class methods, and module
functions. We saw that most methods are normal methods, with class methods
being occasionally useful. Static methods are rarely used, since class methods
or module functions are almost always better alternatives.
The built-in repr() method calls an object’s __repr__() special method. Where
possible, eval(repr(x)) == x, and we saw how to support this. When an
eval()-able representation string cannot be produced we use the base class object.__repr__() method to produce a non-eval()-able representation in a standard format.
Type testing using the built-in isinstance() function can provide some efficiency benefits, although object-oriented purists would almost certainly avoid its
use. Accessing base class methods is achieved by calling the built-in super()
function, and is essential to avoid infinite recursion when we need to call a base
class method inside a subclass’s reimplementation of that method.
Generator functions and methods do lazy evaluation, returning (via the yield
expression) each value one at a time on request and raising a StopIteration
when (and if) they run out of values. Generators can be used wherever an
iterator is expected, and for finite generators, all their values can be extracted
into a tuple or list by passing the iterator returned by the generator to tuple()
or list().
The object-oriented approach almost invariably simplifies code compared with
a purely procedural approach. With custom classes we can guarantee that only
valid operations are available (since we implement only appropriate methods),
and that no operation can put an object into an invalid state (e.g., by using
properties to apply validation). Once we start using object orientation our style
of programming is likely to change from being about global data structures
and the global functions that are applied to the data, to creating classes and
implementing the methods that are applicable to them. Object orientation
makes it possible to package up data and those methods that make sense for
the data. This helps us avoid mixing up all our data and functions together, and
makes it easier to produce maintainable programs since functionality is kept
separated out into individual classes.
The first two exercises involve modifying classes we covered in this chapter,
and the last two exercises involve creating new classes from scratch.
1. Modify the Point class (from or, to support the
following operations, where p, q, and r are Points and n is a number:
= q + r
+= q
= q - r
-= q
= q * n
*= n
= q / n
/= n
= q // n
//= n
The in-place methods are all four lines long, including the def line, and
the other methods are each just two lines long, including the def line,
and of course they are all very similar and quite simple. With a minimal
description and a doctest for each it adds up to around one hundred thirty
new lines. A model solution is provided in; the same code is
also in
2. Modify the class to provide a resize(width, height) method. If the
new width or height is smaller than the current value, any colors outside
the new boundaries must be deleted. If either width or height is None then
use the existing width or height. At the end, make sure you regenerate
the self.__colors set. Return a Boolean to indicate whether a change
was made or not. The method can be implemented in fewer than 20 lines
(fewer than 35 including a docstring with a simple doctest). A solution is
provided in
3. Implement a Transaction class that takes an amount, a date, a currency (default “USD”—U.S. dollars), a USD conversion rate (default 1),
and a description (default None). All of the data attributes must be private. Provide the following read-only properties: amount, date, currency, usd_conversion_rate, description, and usd (calculated from amount *
usd_conversion_rate). This class can be implemented in about sixty lines
including some simple doctests. A model solution for this exercise (and the
next one) is in file
4. Implement an Account class that holds an account number, an account
name, and a list of Transactions. The number should be a read-only
Chapter 6. Object-Oriented Programming
erty; the name should be a read-write property with an assertion to ensure
that the name is at least four characters long. The class should support
the built-in len() function (returning the number of transactions), and
should provide two calculated read-only properties: balance which should
return the account’s balance in USD and all_usd which should return
True if all the transactions are in USD and False otherwise. Three other
methods should be provided: apply() to apply (add) a transaction, save(),
and load(). The save() and load() methods should use a binary pickle
with the filename being the account number with extension .acc; they
should save and load the account number, the name, and all the transactions. This class can be implemented in about ninety lines with some
simple doctests that include saving and loading—use code such as name
= os.path.join(tempfile.gettempdir(), account_name) to provide a suitable
temporary filename, and make sure you delete the temporary file after the
tests have finished. A model solution is in file
● Writing and Reading Binary Data
● Writing and Parsing Text Files
● Writing and Parsing XML Files
● Random Access Binary Files
File Handling
Most programs need to save and load information, such as data or state
information, to and from files. Python provides many different ways of doing
this. We already briefly discussed handling text files in Chapter 3 and pickles
in the preceding chapter. In this chapter we will cover file handling in much
more depth.
All the techniques presented in this chapter are platform-independent. This
means that a file saved using one of the example programs on one operating
system/processor architecture combination can be loaded by the same program
on a machine with a different operating system/processor architecture combination. And this can be true of your programs too if you use the same techniques as the example programs.
The chapter’s first three sections cover the common case of saving and loading
an entire data collection to and from disk. The first section shows how to do this
using binary file formats, with one subsection using (optionally compressed)
pickles, and the other subsection showing how to do the work manually. The
second section shows how to handle text files. Writing text is easy, but reading
it back can be tricky if we need to handle nontextual data such as numbers
and dates. We show two approaches to parsing text, doing it manually and
using regular expressions. The third section shows how to read and write XML
files. This section covers writing and parsing using element trees, writing and
parsing using the DOM (Document Object Model), and writing manually and
parsing using SAX (Simple API for XML).
The fourth section shows how to handle random access binary files. This is
useful when each data item is the same size and where we have more items
than we want in (or can fit into) memory.
Which is the best file format to use for holding entire collections—binary, text,
or XML? Which is the best way to handle each format? These questions are too
context-dependent to have a single definitive answer, especially since there are
Chapter 7. File Handling
Data Type
Minimum length 8 and no whitespace
Nonempty and no newlines
Nonempty and no newlines
Nonempty and no newlines
Range 0.0 to 100.0
Positive and nonzero
Figure 7.1 Aircraft incident record
pros and cons for each format and for each way of handling them. We show all
of them to help you make an informed decision on a case-by-case basis.
Binary formats are usually very fast to save and load and they can be very
compact. Binary data doesn’t need parsing since each data type is stored using
its natural representation. Binary data is not human readable or editable, and
without knowing the format in detail it is not possible to create separate tools
to work with binary data.
Text formats are human readable and editable, and this can make text files
easier to process with separate tools or to change using a text editor. Text
formats can be tricky to parse and it is not always easy to give good error
messages if a text file’s format is broken (e.g., by careless editing).
XML formats are human readable and editable, although they tend to be
verbose and create large files. Like text formats, XML formats can be processed
using separate tools. Parsing XML is straightforward (providing we use an
XML parser rather than do it manually), and some parsers have good error
reporting. XML parsers can be slow, so reading very large XML files can take
a lot more time than reading an equivalent binary or text file. XML includes
metadata such as the character encoding (either implicitly or explicitly) that
is not often provided in text files, and this can make XML more portable than
text files.
Text formats are usually the most convenient for end-users, but sometimes
performance issues are such that a binary format is the only reasonable
choice. However, it is always useful to provide import/export for XML since
this makes it possible to process the file format with third-party tools without
preventing the most optimal text or binary format being used by the program
for normal processing.
Chapter 7. File Handling
Reader + Writer
Lines of Code
Lines of Code
Output File
Size (~KB)
Pickle (gzip compressed)
20 + 16
20 + 16
Manual (gzip compressed)
60 + 34
60 + 34
Plain text
Regex reader/manual writer
39 + 28
Plain text
53 + 28
Element tree
37 + 27
44 + 36
SAX reader/manual writer
55 + 37
Figure 7.2 Aircraft incident file format reader/writer comparison
This chapter’s first three sections all use the same data collection: a set of aircraft incident records. Figure 7.1 shows the names, data types, and validation
constraints that apply to aircraft incident records. It doesn’t really matter
what data we are processing. The important thing is that we learn to process
the fundamental data types including strings, integers, floating-point numbers,
Booleans, and dates, since if we can handle these we can handle any other kind
of data.
By using the same set of aircraft incident data for binary, text, and XML
formats, it makes it possible to compare and contrast the different formats and
the code necessary for handling them. Figure 7.2 shows the number of lines of
code for reading and writing each format, and the totals.
The file sizes are approximate and based on a particular sample of 596 aircraft
incident records.★ Compressed binary file sizes for the same data saved under
different filenames may vary by a few bytes since the filename is included in
the compressed data and filename lengths vary. Similarly, the XML file sizes
vary slightly since some XML writers use entities (&quot; for " and &apos; for ')
for quotes inside text data, and others don’t.
The first three sections all quote code from the same program: This program is used to read aircraft incident data in one format and
to write it in another format. Here is the program’s console help text. (We have
reformatted the output slightly to fit the book’s page width.)
Usage: [options] infile outfile
The data we used is based on real aircraft incident data available from the FAA (U.S. government’s
Federal Aviation Administration,
Chapter 7. File Handling
Reads aircraft incident data from infile and writes the data to
outfile. The data formats used depend on the file extensions:
.aix is XML, .ait is text (UTF-8 encoding), .aib is binary,
.aip is pickle, and .html is HTML (only allowed for the outfile).
All formats are platform-independent.
-h, --help
show this help message and exit
-f, --force
write the outfile even if it exists [default: off]
-v, --verbose
report results [default: off]
-r READER, --reader=READER
reader (XML): 'dom', 'd', 'etree', 'e', 'sax', 's'
reader (text): 'manual', 'm', 'regex', 'r'
[default: etree for XML, manual for text]
-w WRITER, --writer=WRITER
writer (XML): 'dom', 'd', 'etree', 'e',
'manual', 'm' [default: manual]
-z, --compress compress .aib/.aip outfile [default: off]
-t, --test
execute doctests and exit (use with -v for verbose)
The options are more complex than would normally be required since an
end-user will not care which reader or writer we use for any particular format.
In a more realistic version of the program the reader and writer options would
not exist and we would implement just one reader and one writer for each
format. Similarly, the test option exists to help us test the code and would not
be present in a production version.
The program defines one custom exception:
class IncidentError(Exception): pass
Aircraft incidents are held as Incident objects. Here is the class line and
the initializer:
class Incident:
def __init__(self, report_id, date, airport, aircraft_id,
aircraft_type, pilot_percent_hours_on_type,
pilot_total_hours, midair, narrative=""):
assert len(report_id) >= 8 and len(report_id.split()) == 1, \
"invalid report ID"
self.__report_id = report_id = date
self.airport = airport
self.aircraft_id = aircraft_id
self.aircraft_type = aircraft_type
self.pilot_percent_hours_on_type = pilot_percent_hours_on_type
Chapter 7. File Handling
self.pilot_total_hours = pilot_total_hours
self.midair = midair
self.narrative = narrative
The report ID is validated when the Incident is created and is available as
the read-only report_id property. All the other data attributes are read/write
properties. For example, here is the date property’s code:
def date(self):
return self.__date
def date(self, date):
assert isinstance(date,, "invalid date"
self.__date = date
All the other properties follow the same pattern, differing only in the details
of their assertions, so we won’t reproduce them here. Since we have used
assertions, the program will fail if an attempt is made to create an Incident
with invalid data, or to set one of an existing incident’s read/write properties to
an invalid value. We have chosen this uncompromising approach because we
want to be sure that the data we save and load is always valid, and if it isn’t we
want the program to terminate and complain rather than silently continue.
The collection of incidents is held as an IncidentCollection. This class is a dict
subclass, so we get a lot of functionality, such as support for the item access
operator ([]) to get, set, and delete incidents, by inheritance. Here is the class
line and a few of the class’s methods:
class IncidentCollection(dict):
def values(self):
for report_id in self.keys():
yield self[report_id]
def items(self):
for report_id in self.keys():
yield (report_id, self[report_id])
def __iter__(self):
for report_id in sorted(super().keys()):
yield report_id
keys = __iter__
We have not needed to reimplement the initializer since dict.__init__() is
sufficient. The keys are report IDs and the values are Incidents. We have
reimplemented the values(), items(), and keys() methods so that their iterators
Writing and Reading Binary Data
The Bytes and Bytearray Data Types
Python provides two data types for handling raw bytes: bytes which is immutable, and bytearray which is mutable. Both types hold a sequence of zero
or more 8-bit unsigned integers (bytes) with each byte in the range 0…255.
77 ➤
Both types are very similar to strings and provide many of the same
methods, including support for slicing. In addition, bytearrays also provide
some mutating list-like methods. All their methods are listed in Tables 7.1
(➤ 299) and 7.2 (➤ 300).
Whereas a slice of a bytes or bytearray returns an object of the same type,
accessing a single byte using the item access operator ([]) returns an
int—the value of the specified byte. For example:
word = b"Animal"
x = b"A"
word[0] == x
# returns: False
word[:1] == x
# returns: True
word[0] == x[0] # returns: True
# word[0] == 65;
x == b"A"
# word[:1] == b"A"; x == b"A"
# word[0] == 65;
x[0] == 65
Here are some other bytes and bytearray examples:
data = b"5 Hills \x35\x20\x48\x69\x6C\x6C\x73"
# returns: b'5 HILLS 5 HILLS'
data.replace(b"ill", b"at")
# returns: b'5 Hats 5 Hats'
bytes.fromhex("35 20 48 69 6C 6C 73") # returns: b'5 Hills'
# returns: b'5 Hills'
data = bytearray(data)
# data is now a bytearray
# returns: 72 (ord("H"))
data.insert(10, ord("B"))
# data == b'5 Hills 5 Bills'
Methods that make sense only for strings, such as bytes.upper(), assume
that the bytes are encoded using ASCII. The bytes.fromhex() class method
ignores whitespace and interprets each two-digit substring as a hexadecimal
number, so "35" is taken to be a byte of value 0x35, and so on.
pickle.dump(self, fh, pickle.HIGHEST_PROTOCOL)
return True
except (EnvironmentError, pickle.PicklingError) as err:
print("{0}: export error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
if fh is not None:
Chapter 7. File Handling
If compression has been requested, we use the gzip module’s
function to open the file; otherwise, we use the built-in open() function. We
must use “write binary” mode ("wb") when pickling data in binary format. In
Python 3.0 and 3.1, pickle.HIGHEST_PROTOCOL is protocol 3, a compact binary
pickle format. This is the best protocol to use for data shared among Python 3
For error handling we have chosen to report errors to the user as soon as they
occur, and to return a Boolean to the caller indicating success or failure. And
we have used a finally block to ensure that the file is closed at the end, whether
there was an error or not. In Chapter 8 we will use a more compact idiom to
ensure that files are closed that avoids the need for a finally block.
➤ 369
This code is very similar to what we saw in the preceding chapter, but there is
one subtle point to note. The pickled data is self, a dict. But the dictionary’s
values are Incident objects, that is, objects of a custom class. The pickle module
is smart enough to be able to save objects of most custom classes without us
needing to intervene.
In general, Booleans, numbers, and strings can be pickled, as can instances of
classes including custom classes, providing their private __dict__ is picklable.
In addition, any built-in collection types (tuples, lists, sets, dictionaries) can
be pickled, providing they contain only picklable objects (including collection
types, so recursive structures are supported). It is also possible to pickle other
kinds of objects or instances of custom classes that can’t normally be pickled
(e.g., because they have a nonpicklable attribute), either by giving some help
to the pickle module or by implementing custom pickle and unpickle functions.
All the relevant details are provided in the pickle module’s online documentation.
➤ 363
To read back the pickled data we need to distinguish between a compressed and
an uncompressed pickle. Any file that is compressed using gzip compression
begins with a particular magic number. A magic number is a sequence of one
or more bytes at the beginning of a file that is used to indicate the file’s type.
For gzip files the magic number is the two bytes 0x1F 0x8B, which we store in a
bytes variable:
GZIP_MAGIC = b"\x1F\x8B"
For more about the bytes data type, see the sidebar “The Bytes and Bytearray
Data Types” (293 ➤), and Tables 7.1, 7.2, and 7.3 (➤ 299–301), which list
their methods.
Here is the code for reading an incidents pickle file:
Protocol 3 is Python 3-specific. If we want pickles that are readable and writable by both Python 2
and Python 3 programs, we must use protocol 2 instead. Note, though, that protocol 2 files written
by Python 3.1 can be read by Python 3.1 and Python 2.x, but not by Python 3.0!
Writing and Reading Binary Data
def import_pickle(self, filename):
fh = None
fh = open(filename, "rb")
magic =
if magic == GZIP_MAGIC:
fh =, "rb")
return True
except (EnvironmentError, pickle.UnpicklingError) as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
if fh is not None:
We don’t know whether the given file is compressed. In either case we begin
by opening the file in “read binary” mode, and then we read the first two bytes.
If these bytes are the same as the gzip magic number we close the file and
create a new file object using the function. And if the file is not
compressed we use the file object returned by open(), calling its seek() method
to restore the file pointer to the beginning so that the next read (made inside
the pickle.load() function) will be from the start.
We can’t assign to self since that would wipe out the IncidentCollection object
that is in use, so instead we clear all the incidents to make the dictionary empty
and then use dict.update() to populate the dictionary with all the incidents
from the IncidentCollection dictionary loaded from the pickle.
Note that it does not matter whether the processor’s byte ordering is big- or
little-endian, because for the magic number we read individual bytes, and for
the data the pickle module handles endianness for us.
Raw Binary Data with Optional Compression
Writing our own code to handle raw binary data gives us complete control
over our file format. It should also be safer than using pickles, since maliciously invalid data will be handled by our code rather than executed by the
Chapter 7. File Handling
When creating custom binary file formats it is wise to create a magic number
to identify your file type, and a version number to identify the version of the
file format in use. Here are the definitions used in the
MAGIC = b"AIB\x00"
FORMAT_VERSION = b"\x00\x01"
We have used four bytes for the magic number and two for the version.
Endianness is not an issue because these will be written as individual bytes,
not as the byte representations of integers, so they will always be the same on
any processor architecture.
To write and read raw binary data we must have some means of converting
Python objects to and from suitable binary representations. Most of the functionality we need is provided by the struct module, briefly described in the sidebar “The Struct Module” (➤ 297), and by the bytes and bytearray data types,
briefly described in the sidebar “The Bytes and Bytearray Data Types” (293 ➤).
The bytes and bytearray classes’ methods are listed in Tables 7.1 (➤ 299) and
7.2 (➤ 300).
Unfortunately, the struct module can handle strings only of a specified length,
and we need variable length strings for the report and aircraft IDs, as well as
for the airport, the aircraft type, and the narrative texts. To meet this need we
have created a function, pack_string(), which takes a string and returns a bytes
object which contains two components: The first is an integer length count, and
the second is a sequence of length count UTF-8 encoded bytes representing the
string’s text.
Since the only place the pack_string() function is needed is inside the export_binary() function, we have put the definition of pack_string() inside the
export_binary() function. This means that pack_string() is not visible outside
the export_binary() function, and makes clear that it is just a local helper function. Here is the start of the export_binary() function, and the complete nested
pack_string() function:
def export_binary(self, filename, compress=False):
def pack_string(string):
data = string.encode("utf8")
format = "<H{0}s".format(len(data))
return struct.pack(format, len(data), data)
91 ➤
The str.encode() method returns a bytes object with the string encoded according to the specified encoding. UTF-8 is a very convenient encoding because it
can represent any Unicode character and is especially compact when representing ASCII characters (just one byte each). The format variable is set to hold
a struct format based on the string’s length. For example, given the string
➤ 351
Writing and Reading Binary Data
The Struct Module
The struct module provides struct.pack(), struct.unpack(), and some other
functions, and the struct.Struct() class. The struct.pack() function takes
a struct format string and one or more values and returns a bytes object
that holds all the values represented in accordance with the format. The
struct.unpack() function takes a format and a bytes or bytearray object and
returns a tuple of the values that were originally packed using the format.
For example:
data = struct.pack("<2h", 11, -9)
items = struct.unpack("<2h", data)
# data == b'\x0b\x00\xf7\xff'
# items == (11, -9)
Format strings consist of one or more characters. Most characters represent
a value of a particular type. If we need more than one value of a type we
can either write the character as many times as there are values of the type
("hh"), or precede the character with a count as we have done here ("2h").
Many format characters are described in the struct module’s online documentation, including “b” (8-bit signed integer), “B” (8-bit unsigned integer),
“h” (16-bit signed integer—used in the examples here), “H” (16-bit unsigned
integer), “i” (32-bit signed integer), “I” (32-bit unsigned integer), “q” (64-bit
signed integer), “Q” (64-bit unsigned integer), “f” (32-bit float), “d” (64-bit
float—this corresponds to Python’s float type), “?” (Boolean), “s” (bytes or
bytearray object—byte strings), and many others.
For some data types such as multibyte integers, the processor’s endianness
makes a difference to the byte order. We can force a particular byte order
to be used regardless of the processor architecture by starting the format
string with an endianness character. In this book we always use “<”, which
means little-endian since that’s the native endianness for the widely used
Intel and AMD processors. Big-endian (also called network byte order) is
signified by “>” (or by “!”). If no endianness is specified the machine’s endianness is used. We recommend always specifying the endianness even if it is
the same as the machine being used since doing so keeps the data portable.
The struct.calcsize() function takes a format and returns how many bytes
a struct using the format will occupy. A format can also be stored by creating
a struct.Struct() object giving it the format as its argument, with the size
of the struct.Struct() object given by its size attribute. For example:
TWO_SHORTS = struct.Struct("<2h")
data = TWO_SHORTS.pack(11, -9)
items = TWO_SHORTS.unpack(data)
# data == b'\x0b\x00\xf7\xff'
# items == (11, -9)
In both examples, 11 is 0x000b, but this is transformed into the bytes 0x0b 0x00
because we have used little-endian byte ordering.
Chapter 7. File Handling
“”, the format will be "<H16s" (little-endian byte order, 2-byte
unsigned integer, 16-byte byte string), and the bytes object that is returned will
be b'\x10\'. Conveniently, Python shows bytes objects in a
compact form using printable ASCII characters where possible, and hexadecimal escapes (and some special escapes like \t and \n) otherwise.
The pack_string() function can handle strings of up to 65 535 UTF-8 characters. We could easily switch to using a different kind of integer for the byte
count; for example, a 4-byte signed integer (format “i”) would allow for strings
of up to 231-1 (more than 2 billion) characters.
The struct module does provide a similar built-in format, “p”, that stores a single byte as a character count followed by up to 255 characters. For packing,
the code using “p” format is slightly simpler than doing all the work ourselves.
But “p” format restricts us to a maximum of 255 UTF-8 characters and provides almost no benefit when unpacking. (For the sake of comparison, versions
of pack_string() and unpack_string() that use “p” format are included in the source file.)
We can now turn our attention to the rest of the code in the export_binary()
fh = None
if compress:
fh =, "wb")
fh = open(filename, "wb")
for incident in self.values():
data = bytearray()
return True
Writing and Reading Binary Data
Table 7.1 Bytes and Bytearray Methods #1
Appends int i (in range 0…255) to bytearray ba
Returns a copy of bytes/bytearray b with the first character capitalized (if it is an ASCII letter),
Returns a copy of b centered in length width padded with
spaces or optionally with the given byte
start, end)
Returns the number of occurrences of bytes/bytearray x in
bytes/ bytearray b (or in the start:end slice of b)
Returns a str object that represents the bytes using the
UTF-8 encoding or using the specified encoding and handling errors according to the optional error argument
91 ➤
start, end)
Returns True if b (or the start:end slice of b) ends with
bytes/ bytearray x or with any of the bytes/bytearrays in
tuple x; otherwise, returns False
Returns a copy of bytes/bytearray b with tabs replaced
with spaces in multiples of 8 or of size if specified
Extends bytearray ba with all the ints in sequence seq; all
the ints must be in the range 0…255
start, end)
Returns the leftmost position of bytes/bytearray x in b
(or in the start:end slice of b) or -1 if not found. Use the
rfind() method to find the rightmost position.
Returns a bytes object with bytes corresponding to the
hexadecimal integers in str h
start, end)
Returns the leftmost position of x in b (or in the start:end
slice of b) or raises ValueError if not found. Use the
rindex() method to find the rightmost position.
ba.insert(p, i)
Inserts integer i (in range 0…255) at position p in ba
Returns True if bytes/bytearray b is nonempty and every
character in b is an ASCII alphanumeric character
Returns True if bytes/bytearray b is nonempty and every
character in b is an ASCII alphabetic character
Returns True if bytes/bytearray b is nonempty and every
character in b is an ASCII digit
Returns True if bytes/bytearray b has at least one lowercaseable ASCII character and all its lowercaseable characters are lowercase
Returns True if bytes/bytearray b is nonempty and every
character in b is an ASCII whitespace character
Chapter 7. File Handling
Table 7.2 Bytes and Bytearray Methods #2
Returns True if b is nonempty and title-cased
Returns True if b has at least one uppercaseable ASCII character and all its uppercaseable characters are uppercase
Returns the concatenation of every bytes/bytearray in sequence seq, with b (which may be empty) between each one
Returns a copy of bytes/bytearray b left-aligned in length
width padded with spaces or optionally with the given byte.
Use the rjust() method to right-align.
Returns an ASCII-lowercased copy of bytes/bytearray b
Returns a tuple of three bytes objects—the part of b before
the leftmost bytes/bytearray sep, sep itself, and the part of
b after sep; or if sep isn’t in b returns b and two empty bytes
objects. Use the rpartition() method to partition on the
rightmost occurrence of sep.
Removes and returns the int at index position p in ba
Removes the first occurrence of int i from bytearray ba
y, n)
Returns a copy of b with every (or a maximum of n if given)
occurrence of bytes/bytearray x replaced with y
Reverses bytearray ba’s bytes in-place
b.split(x, n)
Returns a list of bytes splitting at most n times on x. If n isn’t
given, splits everywhere possible; if x isn’t given, splits on
whitespace. Use rsplit() to split from the right.
Returns the list of lines produced by splitting b on line
terminators, stripping the terminators unless f is True
x, start,
Returns True if bytes/bytearray b (or the start:end slice
of b) starts with bytes/bytearray x or with any of the
bytes/ bytearrays in tuple x; otherwise, returns False
Returns a copy of b with leading and trailing whitespace (or
the bytes in bytes/bytearray x) removed; lstrip() strips only
at the start, and rstrip() strips only at the end
Returns a copy of b with uppercase ASCII characters lowercased and lowercase ASCII characters uppercased
Returns a copy of b where the first ASCII letter of each word
is uppercased and all other ASCII letters are lowercased
bt, d)
Returns a copy of b that has no bytes from d, and where each
other byte is replaced by the byte-th byte from bytes bt
Writing and Reading Binary Data
Table 7.3 Bytes and Bytearray Methods #3
Returns an ASCII-uppercased copy of bytes/bytearray b
Returns a copy of b, which if shorter than w is padded with
leading zeros (0x30 characters) to make it w bytes long
We have omitted the except and finally blocks since they are the same as the
ones shown in the preceding subsection, apart from the particular exceptions
that the except block catches.
We begin by opening the file in “write binary” mode, either a normal file or a
gzip compressed file depending on the compress flag. We then write the 4-byte
magic number that is (hopefully) unique to our program, and the 2-byte version
number.★ Using a version number makes it easier to change the format in the
future—when we read the version number we can use it to determine which
code to use for reading.
Next we iterate over all the incidents, and for each one we create a bytearray.
We add each item of data to the byte array, starting with the variable length
strings. The date.toordinal() method returns a single integer representing
the stored date; the date can be restored by passing this integer to the method. The NumbersStruct is defined earlier in the
program with this statement:
NumbersStruct = struct.Struct("<Idi?")
This format specifies little-endian byte order, an unsigned 32-bit integer (for
the date ordinal), a 64-bit float (for the percentage hours on type), a 32-bit integer (for the total hours flown), and a Boolean (for whether the incident was
midair). The structure of an entire aircraft incident record is shown schematically in Figure 7.3.
Once the bytearray has all the data for one incident, we write it to disk. And
once all the incidents have been written we return True (assuming no error occurred). The finally block ensures that the file is closed just before we return.
Reading back the data is not as straightforward as writing it—for one thing
we have more error checking to do. Also, reading back variable length strings
is slightly tricky. Here is the start of the import_binary() method and the
complete nested unpack_string() function that we use to read back the variable
length strings:
There is no central repository for magic numbers like there is for domain names, so we can never
guarantee uniqueness.
hou t_to
hou t_pe
rs_ rce
on_ nt_
Chapter 7. File Handling
UTF-8 encoded bytes...
Figure 7.3 The structure of a binary aircraft incident record
def import_binary(self, filename):
def unpack_string(fh, eof_is_error=True):
uint16 = struct.Struct("<H")
length_data =
if not length_data:
if eof_is_error:
raise ValueError("missing or corrupt string size")
return None
length = uint16.unpack(length_data)[0]
if length == 0:
return ""
data =
if not data or len(data) != length:
raise ValueError("missing or corrupt string")
format = "<{0}s".format(length)
return struct.unpack(format, data)[0].decode("utf8")
Since each incident record begins with its report ID string, when we attempt to
read this string and we succeed, we are at the start of a new record. But if we
fail, we’ve reached the end of the file and can finish. We set the eof_is_error
flag to False when attempting to read a report ID since if there is no data, it
just means we have finished. For all other strings we accept the default of True
because if any other string has no data, it is an error. (Even an empty string
will be preceded by a 16-bit unsigned integer length.)
We begin by attempting to read the string’s length. If this fails we return None
to signify end of file (if we are attempting to read a new incident), or we raise
a ValueError exception to indicate corrupt or missing data. The struct.unpack()
function and the struct.Struct.unpack() method always return a tuple, even
if it contains only a single value. We unpack the length data and store the
number it represents in the length variable. Now we know how many bytes we
Writing and Reading Binary Data
must read to get the string. If the length is zero we simply return an empty
string. Otherwise, we attempt to read the specified number of bytes. If we
don’t get any data or if the data is not the size we expected (i.e., it is too little),
we raise a ValueError exception.
If we have the right number of bytes we create a suitable format string for the
struct.unpack() function, and we return the string that results from unpacking
the data and decoding the bytes as UTF-8. (In theory, we could replace the
last two lines with return data.decode("utf8"), but we prefer to go through the
unpacking process since it is possible—though unlikely—that the “s” format
performs some transformation on our data which must be reversed when
reading back.)
We will now look at the rest of the import_binary() method, breaking it into two
parts for ease of explanation.
fh = None
fh = open(filename, "rb")
magic =
if magic == GZIP_MAGIC:
fh =, "rb")
magic =
if magic != MAGIC:
raise ValueError("invalid .aib file format")
version =
if version > FORMAT_VERSION:
raise ValueError("unrecognized .aib file version")
The file may or may not be compressed, so we use the same technique that
we used for reading pickles to open the file using or the built-in
open() function.
Once the file is open and we are at the beginning, we read the first four bytes
(len(MAGIC)). If these don’t match our magic number we know that it isn’t a
binary aircraft incident data file and so we raise a ValueError exception. Next
we read in the 2-byte version number. It is at this point that we would use
different reading code depending on the version. Here we just check that the
version isn’t a later one than this program is able to read.
If the magic number is correct and the version is one we can handle, we are
ready to read in the data, so we begin by clearing out all the existing incidents
so that the dictionary is empty.
Chapter 7. File Handling
while True:
report_id = unpack_string(fh, False)
if report_id is None:
data = {}
data["report_id"] = report_id
for name in ("airport", "aircraft_id",
"aircraft_type", "narrative"):
data[name] = unpack_string(fh)
other_data =
numbers = NumbersStruct.unpack(other_data)
data["date"] =[0])
data["pilot_percent_hours_on_type"] = numbers[1]
data["pilot_total_hours"] = numbers[2]
data["midair"] = numbers[3]
incident = Incident(**data)
self[incident.report_id] = incident
return True
The while block loops until we run out of data. We start by trying to get a report
ID. If we get None we’ve reached the end of the file and can break out of the loop.
Otherwise, we create a dictionary called data to hold the data for one incident
and attempt to get the rest of the incident’s data. For the strings we use the
unpack_string() method, and for the other data we read it all in one go using
the NumbersStruct struct. Since we stored the date as an ordinal we must do
the reverse conversion to get a date back. But for the other items, we can just
use the unpacked data—no validation or conversion is required since we wrote
the correct data types in the first place and have read back the same data types
using the format held in the NumbersStruct struct.
If any error occurs, for example, if we fail to unpack all the numbers, an
exception will be raised and will be handled in the except block. (We haven’t
shown the except and finally blocks because they are structurally the same as
those shown in the preceding subsection for the import_pickle() method.)
179 ➤
Toward the end we make use of the convenient mapping unpacking syntax to
create an Incident object which we then store in the incidents dictionary.
Apart from the handling of variable length strings, the struct module makes
it very easy to save and load data in binary format. And for variable length
strings the pack_string() and unpack_string() methods shown here should
serve most purposes perfectly well.
Writing and Parsing Text Files
Writing and Parsing Text Files
Writing text is easy, but reading it back can be problematic, so we need to
choose the structure carefully so that it is not too difficult to parse.★ Figure 7.4
shows an example aircraft incident record in the text format we are going to
use. When we write the incident records to a file we will follow each one with
a blank line, but when we parse the file we will accept zero or more blank lines
between incident records.
Writing Text
Each incident record begins with the report ID enclosed in brackets ([]). This is
followed by all the one-line data items written in key=value form. For the multiline narrative text we precede the text with a start marker (.NARRATIVE_START.)
and follow it with an end marker (.NARRATIVE_END.), and we indent all the text
in between to ensure that no line of text could be confused with a start or end
Figure 7.4 An example text format aircraft incident record
Here is the code for the export_text() function, but excluding the except and
finally blocks since they are the same as ones we have seen before, except for
the exceptions handled:
def export_text(self, filename):
wrapper = textwrap.TextWrapper(initial_indent="
Chapter 14 introduces various parsing techniques, including two third-party open source parsing
modules that make parsing tasks much easier.
Chapter 7. File Handling
fh = None
fh = open(filename, "w", encoding="utf8")
for incident in self.values():
narrative = "\n".join(wrapper.wrap(
return True
The line breaks in the narrative text are not significant, so we can wrap the
text as we like. Normally we would use the textwrap module’s textwrap.wrap()
function, but here we need to both indent and wrap, so we begin by creating a
textwrap.TextWrap object, initialized with the indentation we want to use (four
spaces for the first and subsequent lines). By default, the object will wrap lines
to a width of 70 characters, although we can change this by passing another
keyword argument.
216 ➤
78 ➤
254 ➤
We could have written this using a triple quoted string, but we prefer to put
in the newlines manually. The textwrap.TextWrapper object provides a wrap()
method that takes a string as input, in this case the narrative text, and returns
a list of strings with suitable indentation and each no longer than the wrap
width. We then join this list of lines into a single string using newline as
the separator.
The incident date is held as a object; we have forced str.format()
to use the string representation when writing the date—this very conveniently produces the date in ISO 8601, YYYY-MM-DD format. We have told
str.format() to write the midair bool as an integer—this produces 1 for True
and 0 for False. In general, using str.format() makes writing text very easy because it handles all of Python’s data types (and custom types if we implement
the __str__() or __format__() special method) automatically.
Writing and Parsing Text Files
Parsing Text
The method for reading and parsing text format aircraft incident records is
longer and more involved than the one used for writing. When reading the
file we could be in one of several states. We could be in the middle of reading
narrative lines; we could be at a key=value line; or we could be at a report ID
line at the start of a new incident. We will look at the import_text_manual()
method in five parts.
def import_text_manual(self, filename):
fh = None
fh = open(filename, encoding="utf8")
data = {}
narrative = None
The method begins by opening the file in “read text” mode. Then we clear
the dictionary of incidents and create the data dictionary to hold the data for
a single incident in the same way as we did when reading binary incident
records. The narrative variable is used for two purposes: as a state indicator
and to store the current incident’s narrative text. If narrative is None it means
that we are not currently reading a narrative; but if it is a string (even an
empty one) it means we are in the process of reading narrative lines.
for lino, line in enumerate(fh, start=1):
line = line.rstrip()
if not line and narrative is None:
if narrative is not None:
if line == ".NARRATIVE_END.":
data["narrative"] = textwrap.dedent(
if len(data) != 9:
raise IncidentError("missing data on "
"line {0}".format(lino))
incident = Incident(**data)
self[incident.report_id] = incident
data = {}
narrative = None
narrative += line + "\n"
Since we are reading line by line we can keep track of the current line number
and use this to provide more informative error messages than is possible when
reading binary files. We begin by stripping off any trailing whitespace from
Chapter 7. File Handling
the line, and if this leaves us with an empty line (and providing we are not in
the middle of a narrative), we simply skip to the next line. This means that the
number of blank lines between incidents doesn’t matter, but that we preserve
any blank lines that are in narrative texts.
If the narrative is not None we know that we are in a narrative. If the line is
the narrative end marker we know that we have not only finished reading the
narrative, but also finished reading all the data for the current incident. In
this case we put the narrative text into the data dictionary (having removed
the indentation with the textwrap.dedent() function), and providing we have
the nine pieces of data we need, we create a new incident and store it in the
dictionary. Then we clear the data dictionary and reset the narrative variable
ready for the next record. On the other hand, if the line isn’t the narrative
end marker, we append it to the narrative—including the newline that was
stripped off at the beginning.
elif (not data and line[0] == "["
and line[-1] == "]"):
data["report_id"] = line[1:-1]
If the narrative is None then we are at either a new report ID or are reading
some other data. We could be at a new report ID only if the data dictionary is
empty (because it starts that way and because we clear it after reading each
incident), and if the line begins with [ and ends with ]. If this is the case we
put the report ID into the data dictionary. This means that this elif condition
will not be True again until the data dictionary is next cleared.
elif "=" in line:
key, value = line.split("=", 1)
if key == "date":
data[key] = datetime.datetime.strptime(value,
elif key == "pilot_percent_hours_on_type":
data[key] = float(value)
elif key == "pilot_total_hours":
data[key] = int(value)
elif key == "midair":
data[key] = bool(int(value))
data[key] = value
elif line == ".NARRATIVE_START.":
narrative = ""
raise KeyError("parsing error on line {0}".format(
Writing and Parsing Text Files
If we are not in a narrative and are not reading a new report ID there are only
three more possibilities: We are reading key=value items, we are at a narrative
start marker, or something has gone wrong.
In the case of reading a line of key=value data, we split the line on the first
= character, specifying a maximum of one split—this means that the value
can safely include = characters. All the data read is in the form of Unicode
strings, so for date, numeric, and Boolean data types we must convert the value
string accordingly.
For dates we use the datetime.datetime.strptime() function (“string parse
time”) which takes a format string and returns a datetime.datetime object. We have used a format string that matches the ISO 8601 date format,
and we use to retrieve a object from
the resultant datetime.datetime object, since we want only a date and not a
date/time. We rely on Python’s built-in type functions, float() and int(), for
the numeric conversions. Note, though that, for example, int("4.0") will
raise a ValueError; if we want to be more liberal in accepting integers, we
could use int(float("4.0")), or if we wanted to round rather than truncate,
round(float("4.0")). To get a bool is slightly subtler—for example, bool("0")
returns True (a nonempty string is True), so we must first convert the string to
an int.
Invalid, missing, or out-of-range values will always cause an exception to be
raised. If any of the conversions fail they raise a ValueError exception. And if
any values are out of range an IncidentError exception will be raised when the
data is used to create a corresponding Incident object.
If the line doesn’t contain an = character, we check to see whether we’ve read
the narrative start marker. If we have, we set the narrative variable to be an
empty string. This means that the first if condition will be True for subsequent
lines, at least until the narrative end marker is read.
If none of the if or elif conditions is True then an error has occurred, so in the
final else clause we raise a KeyError exception to signify this.
return True
except (EnvironmentError, ValueError, KeyError,
IncidentError) as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
if fh is not None:
After reading all the lines, we return True to the caller—unless an exception
occurred, in which case the except block catches the exception, prints an error
Chapter 7. File Handling
message for the user, and returns False. And no matter what, if the file was
opened, it is closed at the end.
Parsing Text Using Regular Expressions
Readers unfamiliar with regular expressions (“regexes”) are recommended to
read Chapter 13 before reading this section—or to skip ahead to the following
section (➤ 312), and return here later if desired.
Using regular expressions to parse text files can often produce shorter code
than doing everything by hand as we did in the previous subsection, but it
can be more difficult to provide good error reporting. We will look at the import_text_regex() method in two parts, first looking at the regular expressions
and then at the parsing—but omitting the except and finally blocks since they
have nothing new to teach us.
def import_text_regex(self, filename):
incident_re = re.compile(
key_value_re = re.compile(r"^\s*(?P<key>[^=]+?)\s*=\s*"
r"(?P<value>.+?)\s*$", re.MULTILINE)
67 ➤
The regular expressions are written as raw strings. This saves us from having to double each backslash (writing each \ as \\)—for example, without using raw strings the second regular expression would have to be written as
"^\\s*(?P<key>[^=]+?)\\s*=\\s*(?P<value>.+?)\\s*$". In this book we always
use raw strings for regular expressions.
The first regular expression, incident_re, is used to capture an entire incident record. One effect of this is that any spurious text between records will
not be noticed. This regular expression really has two parts. The first is
\[(?P<id>[^]]+)\](?P<keyvalues>.+?) which matches a [, then matches and captures into the id match group as many non-] characters as it can, then matches a ] (so this gives us the report ID), and then matches as few—but at least
one—of any characters (including newlines because of the re.DOTALL flag), into
the keyvalues match group. The characters matched for the keyvalues match
group are the minimum necessary to take us to the second part of the regular
The second part of the first regular expression is ^\.NARRATIVE_START\.$
(?P<narrative>.*?)^\.NARRATIVE_END\.$ and this matches the literal text .NARRATIVE_START., then as few characters as possible which are captured into the
narrative match group, and then the literal text .NARRATIVE_END., at the end of
Writing and Parsing Text Files
the incident record. The re.MULTILINE flag means that in this regular expression ^ matches at the start of every line (rather than just at the start of the
string), and $ matches at the end of every line (rather than just at the end of
the string), so the narrative start and end markers are matched only at the
start of lines.
The second regular expression, key_value_re, is used to capture key=value lines,
and it matches at the start of every line in the text it is given to match against,
where the line begins with any amount of whitespace (including none), followed by non-= characters which are captured into the key match group, followed by an = character, followed by all the remaining characters in the line
(excluding any leading or trailing whitespace), and captures them into the value match group.
The fundamental logic used to parse the file is the same as we used for the
manual text parser that we covered in the previous subsection, only this time
we extract incident records and incident data within those records using
regular expressions rather than reading line by line.
fh = None
fh = open(filename, encoding="utf8")
for incident_match in incident_re.finditer(
data = {}
data["report_id"] ="id")
data["narrative"] = textwrap.dedent("narrative")).strip()
keyvalues ="keyvalues")
for match in key_value_re.finditer(keyvalues):
data["key")] ="value")
data["date"] = datetime.datetime.strptime(
data["date"], "%Y-%m-%d").date()
data["pilot_percent_hours_on_type"] = (
data["pilot_total_hours"] = int(
data["midair"] = bool(int(data["midair"]))
if len(data) != 9:
raise IncidentError("missing data")
incident = Incident(**data)
self[incident.report_id] = incident
return True
The re.finditer() method returns an iterator which produces each nonoverlapping match in turn. We create a data dictionary to hold one incident’s data
as we have done before, but this time we get the report ID and narrative text
Chapter 7. File Handling
from each match of the incident_re regular expression. We then extract all
the key=value strings in one go using the keyvalues match group, and apply
the key_value_re regular expression’s re.finditer() method to iterate over each
individual key=value line. For each (key, value) pair found, we put them in the
data dictionary—so all the values go in as strings. Then, for those values which
should not be strings, we replace them with a value of the appropriate type
using the same string conversions that we used when parsing the text manually.
We have added a check to ensure that the data dictionary has nine items because if an incident record is corrupt, the key_value.finditer() iterator might
match too many or too few key=value lines. The end is the same as before—we
create a new Incident object and put it in the incidents dictionary, then return
True. If anything went wrong, the except suite will issue a suitable error message and return False, and the finally suite will close the file.
One of the things that makes both the manual and the regular expression
text parsers as short and straightforward as they are is Python’s exceptionhandling. The parsers don’t have to check any of the conversions of strings to
dates, numbers, or Booleans, and they don’t have to do any range checking (the
Incident class does that). If any of these things fail, an exception will be raised,
and we handle all the exceptions neatly in one place at the end. Another benefit of using exception-handling rather than explicit checking is that the code
scales well—even if the record format changes to include more data items, the
error handling code doesn’t need to grow any larger.
Writing and Parsing XML Files
Some programs use an XML file format for all the data they handle, whereas
others use XML as a convenient import/export format. The ability to import
and export XML is useful and is always worth considering even if a program’s
main format is a text or binary format.
Out of the box, Python offers three ways of writing XML files: manually writing the XML, creating an element tree and using its write() method, and creating a DOM and using its write() method. Similarly, for reading and parsing
XML files there are four out-of-the-box approaches that can be used: manually
reading and parsing the XML (not recommended and not covered here—it can
be quite difficult to handle some of the more obscure and advanced details correctly), or using an element tree, DOM, or SAX parser. In addition, there are
also third-party XML libraries available, such as the lxml library mentioned in
Chapter 5 (227 ➤), that are well worth investigating.
The aircraft incident XML format is shown in Figure 7.5. In this section we will
show how to write this format manually and how to write it using an element
tree and a DOM, as well as how to read and parse this format using the element
Writing and Parsing XML Files
<?xml version="1.0" encoding="UTF-8"?>
<incident report_id="20070222008099G" date="2007-02-22"
aircraft_id="80342" aircraft_type="CE-172-M"
pilot_total_hours="440" midair="0">
Figure 7.5 An example XML format aircraft incident record in context
tree, DOM, and SAX parsers. If you don’t care which approach is used to
read or write the XML, you could just read the Element Trees subsection that
follows, and then skip to the chapter’s final section (Random Access Binary
Files; ➤ 324).
Element Trees
Writing the data using an element tree is done in two phases: First an element
tree representing the data must be created, and second the tree must be
written to a file. Some programs might use the element tree as their data
structure, in which case they already have the tree and can simply write out
the data. We will look at the export_xml_etree() method in two parts:
def export_xml_etree(self, filename):
root = xml.etree.ElementTree.Element("incidents")
for incident in self.values():
element = xml.etree.ElementTree.Element("incident",
Chapter 7. File Handling
airport = xml.etree.ElementTree.SubElement(element,
airport.text = incident.airport.strip()
narrative = xml.etree.ElementTree.SubElement(element,
narrative.text = incident.narrative.strip()
tree = xml.etree.ElementTree.ElementTree(root)
We begin by creating the root element (<incidents>). Then we iterate over all
the incident records. For each one we create an element (<incident>) to hold the
data for the incident, and use keyword arguments to provide the attributes. All
the attributes must be text, so we convert the date, numeric, and Boolean data
items accordingly. We don’t have to worry about escaping “&”, “<”, and “>” (or
about quotes in attribute values), since the element tree module (and the DOM
and SAX modules) automatically take care of these details.
Each <incident> has two subelements, one holding the airport name and the
other the narrative text. When subelements are created we must provide the
parent element and the tag name. An element’s read/write text attribute is
used to hold its text.
Once the <incident> has been created with all its attributes and its <airport>
and <narrative> subelements, we add the incident to the hierarchy’s root (<incidents>) element. At the end we have a hierarchy of elements that contains all
the incident record data, which we then trivially convert into an element tree.
tree.write(filename, "UTF-8")
except EnvironmentError as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
return True
Writing the XML to represent an entire element tree is simply a matter of
telling the tree to write itself to the given file using the given encoding.
Up to now when we have specified an encoding we have almost always used
the string "utf8". This works fine for Python’s built-in open() function which
can accept a wide range of encodings and a variety of names for them, such as
“UTF-8”, “UTF8”, “utf-8”, and “utf8”. But for XML files the encoding name can
be only one of the official names, so "utf8" is not acceptable, which is why we
have used "UTF-8".★
See and for information about XML encodings.
Writing and Parsing XML Files
Reading an XML file using an element tree is not much harder than writing
one. Again there are two phases: First we read and parse the XML file, and
then we traverse the resultant element tree to read off the data to populate
the incidents dictionary. Again this second phase is not necessary if the element tree itself is being used as the in-memory data store. Here is the import_xml_etree() method, split into two parts.
def import_xml_etree(self, filename):
tree = xml.etree.ElementTree.parse(filename)
except (EnvironmentError,
xml.parsers.expat.ExpatError) as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
By default, the element tree parser uses the expat XML parser under the hood
which is why we must be ready to catch expat exceptions.
for element in tree.findall("incident"):
data = {}
for attribute in ("report_id", "date", "aircraft_id",
"pilot_total_hours", "midair"):
data[attribute] = element.get(attribute)
data["date"] = datetime.datetime.strptime(
data["date"], "%Y-%m-%d").date()
data["pilot_percent_hours_on_type"] = (
data["pilot_total_hours"] = int(
data["midair"] = bool(int(data["midair"]))
data["airport"] = element.find("airport").text.strip()
narrative = element.find("narrative").text
data["narrative"] = (narrative.strip()
if narrative is not None else "")
incident = Incident(**data)
self[incident.report_id] = incident
except (ValueError, LookupError, IncidentError) as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
return True
Chapter 7. File Handling
Once we have the element tree we can iterate over every <incident> using
the xml.etree.ElementTree.findall() method. Each incident is returned as an
xml.etree.Element object. We use the same technique for handling the element
attributes as we did in the previous section’s import_text_regex() method—first
we store all the values in the data dictionary, and then we convert those values which are dates, numbers, or Booleans to the correct type. For the airport
and narrative subelements we use the xml.etree.Element.find() method to
find them and read their text attributes. If a text element has no text its text
attribute will be None, so we must account for this when reading the narrative
text element since it might be empty. In all cases, the attribute values and
text returned to us do not contain XML escapes since they are automatically
As with all the XML parsers used to process aircraft incident data, an exception will occur if the aircraft or narrative element is missing, or if one of the
attributes is missing, or if one of the conversions fails, or if any of the numeric
data is out of range—this ensures that invalid data will cause parsing to stop
and for an error message to be output. The code at the end for creating and storing incidents and for handling exceptions is the same as we have seen before.
DOM (Document Object Model)
The DOM is a standard API for representing and manipulating an XML
document in memory. The code for creating a DOM and writing it to a file,
and for parsing an XML file using a DOM, is structurally very similar to the
element tree code, only slightly longer.
We will begin by reviewing the export_xml_dom() method in two parts. This
method works in two phases: First a DOM is created to reflect the incident
data, and then the DOM is written out to a file. Just as with an element tree,
some programs might use the DOM as their data structure, in which case they
can simply write out the data.
def export_xml_dom(self, filename):
dom = xml.dom.minidom.getDOMImplementation()
tree = dom.createDocument(None, "incidents", None)
root = tree.documentElement
for incident in self.values():
element = tree.createElement("incident")
for attribute, value in (
("report_id", incident.report_id),
("aircraft_id", incident.aircraft_id),
("aircraft_type", incident.aircraft_type),
Writing and Parsing XML Files
("midair", str(int(incident.midair)))):
element.setAttribute(attribute, value)
for name, text in (("airport", incident.airport),
("narrative", incident.narrative)):
text_element = tree.createTextNode(text)
name_element = tree.createElement(name)
The method begins by getting a DOM implementation. By default, the implementation is provided by the expat XML parser. The xml.dom.minidom module
provides a simpler and smaller DOM implementation than that provided by
the xml.dom module, although the objects it uses are from the xml.dom module.
Once we have a DOM implementation we can create a document. The first argument to xml.dom.DOMImplementation.createDocument() is the namespace URI
which we don’t need, so we pass None; the second argument is a qualified name
(the tag name for the root element), and the third argument is the document
type, and again we pass None since we don’t have a document type. Having
gotten the tree that represents the document, we retrieve the root element and
then proceed to iterate over all the incidents.
For each incident we create an <incident> element, and for each attribute we
want the incident to have we call setAttribute() with the attribute’s name and
value. Just as with the element tree, we don’t have to worry about escaping
“&”, “<”, and “>” (or about quotes in attribute values). For the airport and narrative text elements we must create a text element to hold the text and a normal element (with the appropriate tag name) as the text element’s parent—we
then add the normal element (and the text element it contains) to the current
incident element. With the incident element complete, we add it to the root.
fh = None
fh = open(filename, "w", encoding="utf8")
tree.writexml(fh, encoding="UTF-8")
return True
314 ➤
We have omitted the except and finally blocks since they are the same as ones
we have already seen. What this piece of code makes clear is the difference
between the encoding string used for the built-in open() function and the
encoding string used for XML files, as we discussed earlier.
Importing an XML document into a DOM is similar to importing into an element tree, but like exporting, it requires more code. We will look at the im-
Chapter 7. File Handling
port_xml_dom() function in three parts, starting with the def line and the nested
get_text() function.
def import_xml_dom(self, filename):
def get_text(node_list):
text = []
for node in node_list:
if node.nodeType == node.TEXT_NODE:
return "".join(text).strip()
The get_text() function iterates over a list of nodes (e.g., a node’s child nodes),
and for each one that is a text node, it extracts the node’s text and appends it
to a list of texts. At the end the function returns all the text it has gathered as
a single string, with whitespace stripped from both ends.
dom = xml.dom.minidom.parse(filename)
except (EnvironmentError,
xml.parsers.expat.ExpatError) as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
Parsing an XML file into a DOM is easy since the module does all the hard work
for us, but we must be ready to handle expat errors since just like an element
tree, the expat XML parser is the default parser used by the DOM classes
under the hood.
for element in dom.getElementsByTagName("incident"):
data = {}
for attribute in ("report_id", "date", "aircraft_id",
"pilot_total_hours", "midair"):
data[attribute] = element.getAttribute(attribute)
data["date"] = datetime.datetime.strptime(
data["date"], "%Y-%m-%d").date()
data["pilot_percent_hours_on_type"] = (
data["pilot_total_hours"] = int(
data["midair"] = bool(int(data["midair"]))
airport = element.getElementsByTagName("airport")[0]
Writing and Parsing XML Files
data["airport"] = get_text(airport.childNodes)
narrative = element.getElementsByTagName(
data["narrative"] = get_text(narrative.childNodes)
incident = Incident(**data)
self[incident.report_id] = incident
except (ValueError, LookupError, IncidentError) as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
return True
Once the DOM exists we clear the current incidents data and iterate over all
the incident tags. For each one we extract the attributes, and for date, numeric, and Booleans we convert them to the correct types in exactly the same way
as we did when using an element tree. The only really significant difference
between using a DOM and an element tree is in the handling of text nodes.
We use the xml.dom.Element.getElementsByTagName() method to get the child elements with the given tag name—in the cases of <airport> and <narrative> we
know there is always one of each, so we take the first (and only) child element
of each type. Then we use the nested get_text() function to iterate over these
tags’ child nodes to extract their texts.
As usual, if any error occurs we catch the relevant exception, print an error
message for the user, and return False.
The differences in approach between DOM and element tree are not great,
and since they both use the same expat parser under the hood, they’re both
reasonably fast.
Manually Writing XML
Writing a preexisting element tree or DOM as an XML document can be done
with a single method call. But if our data is not already in one of these forms
we must create an element tree or DOM first, in which case it may be more
convenient to simply write out our data directly.
When writing XML files we must make sure that we properly escape text and
attribute values, and that we write a well-formed XML document. Here is the
export_xml_manual() method for writing out the incidents in XML:
def export_xml_manual(self, filename):
fh = None
fh = open(filename, "w", encoding="utf8")
fh.write('<?xml version="1.0" encoding="UTF-8"?>\n')
➤ 351
Chapter 7. File Handling
for incident in self.values():
fh.write('<incident report_id={report_id} '
'date="{!s}" '
'aircraft_id={aircraft_id} '
'aircraft_type={aircraft_type} '
'"{0.pilot_percent_hours_on_type}" '
'pilot_total_hours="{0.pilot_total_hours}" '
incident.narrative.strip()), 70))))
return True
As we have often done in this chapter, we have omitted the except and finally
We write the file using the UTF-8 encoding and must specify this to the built-in
open() function. Strictly speaking, we don’t have to specify the encoding in the
<?xml?> declaration since UTF-8 is the default encoding, but we prefer to be
explicit. We have chosen to quote all the attribute values using double quotes
("), and so for convenience have used single quotes to quote the string we put
the incidents in to avoid the need to escape the quotes.
The sax.saxutils.quoteattr() function is similar to the sax.saxutils.escape()
function we use for XML text in that it properly escapes “&”, “<”, and “>”
characters. In addition, it escapes quotes (if necessary), and returns a string
that has quotes around it ready for use. This is why we have not needed to put
quotes around the report ID and other string attribute values.
The newlines we have inserted and the text wrapping for the narrative are
purely cosmetic. They are designed to make the file easier for humans to read
and edit, but they could just as easily be omitted.
Writing the data in HTML format is not much different from writing XML. The program includes the export_html() function as a simple
Writing and Parsing XML Files
example of this, although we won’t review it here because it doesn’t really show
anything new.
Parsing XML with SAX (Simple API for XML)
Unlike the element tree and DOM, which represent an entire XML document
in memory, SAX parsers work incrementally, which can potentially be both
faster and less memory-hungry. A performance advantage cannot be assumed,
however, especially since both the element tree and DOM use the fast expat
SAX parsers work by announcing “parsing events” when they encounter start
tags, end tags, and other XML elements. To be able to handle those events
that we are interested in we must create a suitable handler class, and provide
certain predefined methods which are called when matching parsing events
take place. The most commonly implemented handler is a content handler,
although it is possible to provide error handlers and other handlers if we want
finer control.
Here is the complete import_xml_sax() method. It is very short because most of
the work is done by the custom IncidentSaxHandler class:
def import_xml_sax(self, filename):
fh = None
handler = IncidentSaxHandler(self)
parser = xml.sax.make_parser()
return True
except (EnvironmentError, ValueError, IncidentError,
xml.sax.SAXParseException) as err:
print("{0}: import error: {1}".format(
os.path.basename(sys.argv[0]), err))
return False
We create the one handler we want to use and then we create a SAX parser and
set its content handler to be the one we have created. Then we give the filename
to the parser’s parse() method and return True if no parsing errors occurred.
We pass self (i.e., this IncidentCollection dict subclass) to the custom IncidentSaxHandler class’s initializer. The handler clears the old incidents away and
then builds up a dictionary of incidents as the file is parsed. Once the parse is
complete the dictionary contains all the incidents that have been read.
Chapter 7. File Handling
class IncidentSaxHandler(xml.sax.handler.ContentHandler):
def __init__(self, incidents):
self.__data = {}
self.__text = ""
self.__incidents = incidents
Custom SAX handler classes must inherit the appropriate base class. This
ensures that for any methods we don’t reimplement (because we are not
interested in the parsing events they handle), the base class version will be
called—and will safely do nothing.
We start by calling the base class’s initializer. This is generally good practice
for all subclasses, although it is not necessary (though harmless) for direct object subclasses. The self.__data dictionary is used to keep one incident’s data,
the self.__text string is used to keep the text of an airport name or of a narrative depending on which we are reading, and the self.__incidents dictionary
is an object reference to the IncidentCollection dictionary which the handler
updates directly. (An alternative design would be to have an independent dictionary inside the handler and to copy it to the IncidentCollection at the end
using dict.clear() and then dict.update().)
def startElement(self, name, attributes):
if name == "incident":
self.__data = {}
for key, value in attributes.items():
if key == "date":
self.__data[key] = datetime.datetime.strptime(
value, "%Y-%m-%d").date()
elif key == "pilot_percent_hours_on_type":
self.__data[key] = float(value)
elif key == "pilot_total_hours":
self.__data[key] = int(value)
elif key == "midair":
self.__data[key] = bool(int(value))
self.__data[key] = value
self.__text = ""
Whenever a start tag and its attributes are read the xml.sax.handler.ContentHandler.startElement() method is called with the tag name and the tag’s
attributes. In the case of an aircraft incidents XML file, the start tags are
<incidents>, which we ignore; <incident>, whose attributes we use to populate
some of the self.__data dictionary; and <airport> and <narrative>, both of
Writing and Parsing XML Files
which we ignore. We always clear the self.__text string when we get a start
tag because no text tags are nested in the aircraft incident XML file format.
We don’t do any exception-handling in the IncidentSaxHandler class. If an exception occurs it will be passed up to the caller, in this case the import_xml_sax()
method, which will catch it and output a suitable error message.
def endElement(self, name):
if name == "incident":
if len(self.__data) != 9:
raise IncidentError("missing data")
incident = Incident(**self.__data)
self.__incidents[incident.report_id] = incident
elif name in frozenset({"airport", "narrative"}):
self.__data[name] = self.__text.strip()
self.__text = ""
When an end tag is read the xml.sax.handler.ContentHandler.endElement()
method is called. If we have reached the end of an incident we should have
all the necessary data, so we create a new Incident object and add it to the
incidents dictionary. If we have reached the end of a text element, we add an
item to the self.__data dictionary with the text that has been accumulated so
far. At the end we clear the self.__text string ready for its next use. (Strictly
speaking, we don’t have to clear it, since we clear it when we get a start tag, but
clearing it could make a difference in some XML formats, for example, where
tags can be nested.)
def characters(self, text):
self.__text += text
When the SAX parser reads text it calls the xml.sax.handler.ContentHandler.characters() method. There is no guarantee that this method will be called
just once with all the text; the text might come in chunks. This is why we
simply use the method to accumulate text, and actually put the text into the
dictionary only when the relevant end tag is reached. (A more efficient implementation would have self.__text be a list with the body of this method being
self.__text.append(text), and with the other methods adapted accordingly.)
Using the SAX API is very different from using element tree or DOM, but
it is just as effective. We can provide other handlers, and can reimplement
additional methods in the content handler to get as much control as we like.
The SAX parser itself does not maintain any representation of the XML
document—this makes SAX ideal for reading XML into our own custom data
collections, and also means that there is no SAX “document” to write out as
XML, so for writing XML we must use one of the approaches described earlier
in this section.
Chapter 7. File Handling
Random Access Binary Files
In the earlier sections we worked on the basis that all of a program’s data
was read into memory in one go, processed, and then all written out in one go.
Modern computers have so much RAM that this is a perfectly viable approach,
even for large data sets. However, in some situations holding the data on disk
and just reading the bits we need and writing back changes might be a better
solution. The disk-based random access approach is most easily done using
a key–value database (a “DBM”), or a full SQL database—both are covered in
Chapter 12—but in this section we will show how to handle random access files
by hand.
We will first present the BinaryRecordFile.BinaryRecordFile class. Instances
of this class represent a generic readable/writable binary file, structured as a
sequence of fixed length records. We will then look at the BikeStock.BikeStock
class which holds a collection of BikeStock.Bike objects as records in a BinaryRecordFile.BinaryRecordFile to see how to make use of binary random access files.
A Generic BinaryRecordFile Class
The BinaryRecordFile.BinaryRecordFile class’s API is similar to a list in that we
can get/set/delete a record at a given index position. When a record is deleted, it
is simply marked “deleted”; this saves having to move all the records that follow
it up to fill the gap, and also means that after a deletion all the original index
positions remain valid. Another benefit is that a record can be undeleted simply by unmarking it. The price we pay for this is that deleting records doesn’t
save any disk space. We will solve this by providing methods to “compact” the
file, eliminating deleted records (and invalidating index positions).
Before reviewing the implementation, let’s look at some basic usage:
Contact = struct.Struct("<15si")
contacts = BinaryRecordFile.BinaryRecordFile(filename, Contact.size)
Here we create a struct (little-endian byte order, a 15-byte byte string, and
a 4-byte signed integer) that we will use to represent each record. Then we
create a BinaryRecordFile.BinaryRecordFile instance with a filename and with
a record size to match the struct we are using. If the file exists it will be opened
with its contents left intact; otherwise, it will be created. In either case it will
be opened in binary read/write mode, and once open, we can write data to it:
contacts[4] = Contact.pack("Abe Baker".encode("utf8"), 762)
contacts[5] = Contact.pack("Cindy Dove".encode("utf8"), 987)
Random Access Binary Files
Table 7.4 File Object Attributes and Methods #1
Closes file object f and sets attribute f.closed to True
Returns True if the file is closed
The encoding used for bytes ↔ str conversions
Returns the underlying file’s file descriptor. (Available only
for file objects that have file descriptors.)
Flushes the file object f
Returns True if the file object is associated with a console.
(Available only for file objects that refer to actual files.)
The mode file object f was opened with
File object f’s filename (if it has one)
The kinds of newline strings encountered in text file f
Returns the next line from file object f. In most cases, this
method is used implicitly, for example, for line in f.
Returns n bytes without moving the file pointer position
Reads at most count bytes from file object f. If count is not
specified then every byte is read from the current file position to the end. Returns a bytes object when reading in binary mode and a str when reading in text mode. If there
is no more to read (end of file), an empty bytes or str is
Returns True if f was opened for reading
Reads at most len(ba) bytes into bytearray ba and returns
the number of bytes read—this is 0 at end of file. (Available
only in binary mode.)
Reads the next line (or up to count bytes if count is specified
and reached before the \n character), including the \n
Reads all the lines to the end of the file and returns them as
a list. If sizehint is given, then reads approximately up to
sizehint bytes if the underlying file object supports this.
Moves the file pointer position (where the next read or write
will take place) to the given offset if whence is not given or is
os.SEEK_SET. Moves the file pointer to the given offset (which
may be negative) relative to the current position if whence
is os.SEEK_CUR or relative to the end if whence is os.SEEK_END.
Writes are always done at the end in append "a" mode no
matter where the file pointer is. In text mode only the return value of tell() method calls should be used as offsets.
Chapter 7. File Handling
Table 7.5 File Object Attributes and Methods #2
Returns True if f supports random access
Returns the current file pointer position relative to the start
of the file
Truncates the file to the current file pointer position, or to
the given size if size is specified
Returns True if f was opened for writing
Writes bytes/bytearray object s to the file if opened in binary
mode or a str object s to the file if opened in text mode
Writes the sequence of objects (strings for text files, byte
strings for binary files) to the file
We can treat the file like a list using the item access operator ([]); here we
assign two byte strings (bytes objects, each containing an encoded string and
an integer) at two record index positions in the file. These assignments will
overwrite any existing content; and if the file doesn’t already have six records,
the earlier records will be created with every byte set to 0x00.
contact_data = Contact.unpack(contacts[5])
contact_data[0].decode("utf8").rstrip(chr(0)) # returns: 'Cindy Dove'
Since the string “Cindy Dove” is shorter than the 15 UTF-8 characters
in the struct, when it is packed it is padded with 0x00 bytes at the end. So
when we retrieve the record, the contact_data will hold the 2-tuple (b'Cindy
Dove\x00\x00\x00\x00\x00', 987). To get the name, we must decode the UTF-8 to
produce a Unicode string, and strip off the 0x00 padding bytes.
Now that we’ve had a glimpse of the class in action, we are ready to review
the code. The BinaryRecordFile.BinaryRecordFile class is in file After the usual preliminaries the file begins with the definitions of a
couple of private byte values:
_DELETED = b"\x01"
_OKAY = b"\x02"
Each record starts with a “state” byte which is either _DELETED or _OKAY (or
b"\x00" in the case of blank records).
Here is the class line and the initializer:
class BinaryRecordFile:
def __init__(self, filename, record_size, auto_flush=True):
self.__record_size = record_size + 1
Random Access Binary Files
mode = "w+b" if not os.path.exists(filename) else "r+b"
self.__fh = open(filename, mode)
self.auto_flush = auto_flush
There are two different record sizes. The BinaryRecordFile.record_size is the
one set by the user and is the record size from the user’s point of view. The
private BinaryRecordFile.__record_size is the real record size and includes the
state byte.
We are careful not to truncate the file when we open it if it already exists (by
using a mode of "r+b"), and to create it if it does not exist (by using a mode of
"w+b")—the "+" part of the mode string is what signifies reading and writing. If
the BinaryRecordFile.auto_flush Boolean is True, the file is flushed before every
read and after every write.
def record_size(self):
return self.__record_size - 1
def name(self):
def flush(self):
def close(self):
We have made the record size and filename into read-only properties. The
record size we report to the user is the one they requested and matches their
records. The flush and close methods simply delegate to the file object.
def __setitem__(self, index, record):
assert isinstance(record, (bytes, bytearray)), \
"binary data required"
assert len(record) == self.record_size, (
"record must be exactly {0} bytes".format(
self.record_size)) * self.__record_size)
if self.auto_flush:
This method supports the brf[i] = data syntax where brf is a binary record file,
i a record index position, and data a byte string. Notice that the record must
be the same size as the size is specified when the binary record file was created.
Chapter 7. File Handling
If the arguments are okay, we move the file position pointer to the first byte of
the record—notice that here we use the real record size, that is, we account for
the state byte. The seek() method moves the file pointer to an absolute byte
position by default. A second argument can be given to make the movement
relative to the current position or to the end. (The attributes and methods
provided by file objects are listed in Tables 7.4 and 7.5.)
Since the item is being set it obviously hasn’t been deleted, so we write the
_OKAY state byte, and then we write the user’s binary record data. The binary
record file does not know or care about the record structure that is being
used—only that records are of the right size.
We do not check whether the index is in range. If the index is beyond the
end of the file the record will be written in the correct position and every byte
between the previous end of the file and the new record will automatically
be set to b"\x00". Such blank records are neither _OKAY nor _DELETED, so we can
distinguish them when we need to.
def __getitem__(self, index):
state =
if state != _OKAY:
return None
When retrieving a record there are four cases that we must account for: The
record doesn’t exist, that is, the given index is beyond the end; the record is
blank; the record has been deleted; and the record is okay. If the record doesn’t
exist the private __seek_to_index() method will raise an IndexError exception.
Otherwise, it will seek to the byte where the record begins and we can read the
state byte. If the state is not _OKAY the record must either be blank or be deleted, in which case we return None; otherwise, we read and return the record. (Another strategy would be to raise a custom exception for blank or deleted records,
say, BlankRecordError or DeletedRecordError, instead of returning None.)
def __seek_to_index(self, index):
if self.auto_flush:
self.__fh.flush(), os.SEEK_END)
end = self.__fh.tell()
offset = index * self.__record_size
if offset >= end:
raise IndexError("no record at index position {0}".format(
Random Access Binary Files
This is a private supporting method used by some of the other methods to
move the file position pointer to the first byte of the record at the given index
position. We begin by checking to see whether the given index is in range.
We do this by seeking to the end of the file (byte offset of 0 from the end), and
using the tell() method to retrieve the byte position we have seeked to.★ If the
record’s offset (index position × real record size) is at or after the end then the
index is out of range and we raise a suitable exception. Otherwise, we seek to
the offset position ready for the next read or write.
def __delitem__(self, index):
state =
if state != _OKAY:
return * self.__record_size)
if self.auto_flush:
First we move the file position pointer to the right place. If the index is in
range (i.e., if no IndexError exception has occurred), and providing the record
isn’t blank or already deleted, we delete the record by overwriting its state byte
with _DELETED.
def undelete(self, index):
state =
if state == _DELETED: * self.__record_size)
if self.auto_flush:
return True
return False
This method begins by finding the record and reading its state byte. If the
record is deleted we overwrite the state byte with _OKAY and return True to
the caller to indicate success; otherwise (for blank or nondeleted records), we
return False.
def __len__(self):
if self.auto_flush:
self.__fh.flush(), os.SEEK_END)
Both Python 3.0 and 3.1 have the seek constants os.SEEK_SET, os.SEEK_CUR, and os.SEEK_END. For
convenience, Python 3.1 also has these constants in its io module (e.g., io.SEEK_SET).
Chapter 7. File Handling
end = self.__fh.tell()
return end // self.__record_size
This method reports how many records are in the binary record file. It does
this by dividing the end byte position (i.e., how many bytes are in the file) by
the size of a record.
We have now covered all the basic functionality offered by the BinaryRecordFile.BinaryRecordFile class. There is one last matter to consider: compacting
the file to eliminate blank and deleted records. There are essentially two approaches we can take to this. One approach is to overwrite blank or deleted
records with records that have higher record index positions so that there are
no gaps, and truncating the file if there are any blank or deleted records at the
end. The inplace_compact() method does this. The other approach is to copy
the nonblank nondeleted records to a temporary file and then to rename the
temporary to the original. Using a temporary file is particularly convenient if
we also want to make a backup. The compact() method does this.
We will start by looking at the inplace_compact() method, in two parts.
def inplace_compact(self):
index = 0
length = len(self)
while index < length:
state =
if state != _OKAY:
for next in range(index + 1, length):
state =
if state == _OKAY:
self[index] = self[next]
del self[next]
index += 1
We iterate over every record, reading the state of each one in turn. If we find a
blank or deleted record we look for the next nonblank nondeleted record in the
file. If we find one we replace the blank or deleted record with the nonblank
nondeleted one and delete the original nonblank nondeleted one; otherwise,
we break out of the while loop entirely since we have run out of nonblank
nondeleted records.
state =
Random Access Binary Files
if state != _OKAY:
limit = None
for index in range(len(self) - 1, 0, -1):
state =
if state != _OKAY:
limit = index
if limit is not None:
self.__fh.truncate(limit * self.__record_size)
If the first record is blank or deleted, then they must all be blank or deleted
since the previous code moved all nonblank nondeleted records to the beginning of the file and blank and deleted ones to the end. In this case we can simply truncate the file to 0 bytes.
If there is at least one nonblank nondeleted record we iterate from the last
record backward toward the first since we know that blank and deleted records
have been moved to the end. The limit variable is set to the earliest blank or
deleted record (or left as None if there are no blank or deleted records), and the
file is truncated accordingly.
An alternative to doing the compacting in-place is to do it by copying to another
file—this is useful if we want to make a backup, as the compact() method that
we will review next shows.
def compact(self, keep_backup=False):
compactfile = + ".$$$"
backupfile = + ".bak"
fh = open(compactfile, "wb")
while True:
data =
if not data:
if data[:1] == _OKAY:
os.rename(, backupfile)
Chapter 7. File Handling
if not keep_backup:
self.__fh = open(, "r+b")
This method creates two files, a compacted file and a backup copy of the
original file. The compacted file starts out with the same name as the original
but with .$$$ tacked on to the end of the filename, and similarly the backup file
has the original filename with .bak tacked on to the end. We read the existing
file record by record, and for those records that are nonblank and nondeleted
we write them to the compacted file. (Notice that we write the real record, that
is, the state byte plus the user record, each time.)
293 ➤
The line if data[:1] == _OKAY: is quite subtle. Both the data object and the _OKAY
object are of type bytes. We want to compare the first byte of the data object
to the (1 byte) _OKAY object. If we take a slice of a bytes object, we get a bytes
object, but if we take a single byte, say, data[0], we get an int—the byte’s value.
So here we compare the 1 byte slice of data (its first byte, the state byte) with
the 1 byte _OKAY object. (Another way of doing it would be to write if data[0]
== _OKAY[0]: which would compare the two int values.)
At the end we rename the original file as the backup and rename the compacted
file as the original. We then remove the backup if keep_backup is False (the
default). Finally, we open the compacted file (which now has the original
filename), ready to be read or written.
The BinaryRecordFile.BinaryRecordFile class is quite low-level, but it can serve
as the basis of higher-level classes that need random access to files of fixed-size
records, as we will see in the next subsection.
Example: The BikeStock Module’s Classes
The BikeStock module uses a BinaryRecordFile.BinaryRecordFile to provide
a simple stock control class. The stock items are bicycles, each represented
by a BikeStock.Bike instance, and the entire stock of bikes is held in a BikeStock.BikeStock instance. The BikeStock.BikeStock class aggregates a dictionary whose keys are bike IDs and whose values are record index positions, into
a BinaryRecordFile.BinaryRecordFile. Here is a brief example of use to get a feel
for how these classes work:
bicycles = BikeStock.BikeStock(bike_file)
value = 0.0
for bike in bicycles:
value += bike.value
bicycles.increase_stock("GEKKO", 2)
for bike in bicycles:
if bike.identity.startswith("B4U"):
Random Access Binary Files
if not bicycles.increase_stock(bike.identity, 1):
print("stock movement failed for", bike.identity)
This snippet opens a bike stock file and iterates over all the bicycle records it
contains to find the total value (sum of price × quantity) of the bikes held. It
then increases the number of “GEKKO” bikes in stock by two and increments
the stock held for all bikes whose bike ID begins with “B4U” by one. All of these
actions take place on disk, so any other process that reads the bike stock file
will always get the most current data.
Although the BinaryRecordFile.BinaryRecordFile works in terms of indexes,
the BikeStock.BikeStock class works in terms of bike IDs. This is managed by
the BikeStock.BikeStock instance holding a dictionary that relates bike IDs
to indexes.
We will begin by looking at the BikeStock.Bike class’s class line and initializer, then we will look at a few selected BikeStock.BikeStock methods, and finally we will look at the code that provides the bridge between BikeStock.Bike
objects and the binary records used to represent them in a BinaryRecordFile.BinaryRecordFile. (All the code is in the file.)
class Bike:
def __init__(self, identity, name, quantity, price):
assert len(identity) > 3, ("invalid bike identity '{0}'"
self.__identity = identity = name
self.quantity = quantity
self.price = price
All of a bike’s attributes are available as properties—the bike ID (self.__identity) as the read-only Bike.identity property and the others as read/write properties with some assertions for validation. In addition, the Bike.value read-only
property returns the quantity multiplied by the price. (We have not shown the
implementation of the properties since we have seen similar code before.)
The BikeStock.BikeStock class provides its own methods for manipulating bike
objects, and they in turn use the writable bike properties.
class BikeStock:
def __init__(self, filename):
self.__file = BinaryRecordFile.BinaryRecordFile(filename,
self.__index_from_identity = {}
for index in range(len(self.__file)):
record = self.__file[index]
if record is not None:
Chapter 7. File Handling
bike = _bike_from_record(record)
self.__index_from_identity[bike.identity] = index
The BikeStock.BikeStock class is a custom collection class that aggregates a
binary record file (self.__file) and a dictionary (self.__index_from_identity)
whose keys are bike IDs and whose values are record index positions.
Once the file has been opened (and created if it didn’t already exist), we iterate
over its contents (if any). Each bike is retrieved and converted from a bytes
object to a BikeStock.Bike using the private _bike_from_record() function, and
the bike’s identity and index are added to the self.__index_from_identity dictionary.
def append(self, bike):
index = len(self.__file)
self.__file[index] = _record_from_bike(bike)
self.__index_from_identity[bike.identity] = index
Appending a new bike is a matter of finding a suitable index position and
setting the record at that position to the bike’s binary representation. We also
take care to update the self.__index_from_identity dictionary.
def __delitem__(self, identity):
del self.__file[self.__index_from_identity[identity]]
Deleting a bike record is easy; we just find its record index position from its
identity and delete the record at that index position. In the case of the BikeStock.BikeStock class we have not made use of the BinaryRecordFile.BinaryRecordFile’s undeletion capability.
def __getitem__(self, identity):
record = self.__file[self.__index_from_identity[identity]]
return None if record is None else _bike_from_record(record)
Bike records are retrieved by bike ID. If there is no such ID the lookup in the
self.__index_from_identity dictionary will raise a KeyError exception, and if
the record is blank or deleted the BinaryRecordFile.BinaryRecordFile will return
None. But if a record is retrieved we return it as a BikeStock.Bike object.
def __change_stock(self, identity, amount):
index = self.__index_from_identity[identity]
record = self.__file[index]
if record is None:
return False
bike = _bike_from_record(record)
bike.quantity += amount
self.__file[index] = _record_from_bike(bike)
return True
Random Access Binary Files
increase_stock = (lambda self, identity, amount:
self.__change_stock(identity, amount))
decrease_stock = (lambda self, identity, amount:
self.__change_stock(identity, -amount))
The private __change_stock() method provides an implementation for the increase_stock() and decrease_stock() methods. The bike’s index position is
found and the raw binary record is retrieved. Then the data is converted to a
BikeStock.Bike object, the change is applied to the bike, and then the record in
the file is overwritten with the binary representation of the updated bike object. (There is also a __change_bike() method that provides an implementation
for the change_name() and change_price() methods, but none of these are shown
because they are very similar to what’s shown here.)
def __iter__(self):
for index in range(len(self.__file)):
record = self.__file[index]
if record is not None:
yield _bike_from_record(record)
This method ensures that BikeStock.BikeStock objects can be iterated over, just
like a list, with a BikeStock.Bike object returned at each iteration, and skipping
blank and deleted records.
8 × UTF-8 encoded bytes
30 × UTF-8 encoded bytes
Figure 7.6 The logical structure of a bike record file
The private _bike_from_record() and _record_from_bike() functions isolate the
binary representation of the BikeStock.Bike class from the BikeStock.BikeStock
class that holds a collection of bikes. The logical structure of a bike record file
is shown in Figure 7.6. The physical structure is slightly different because each
record is preceded by a state byte.
_BIKE_STRUCT = struct.Struct("<8s30sid")
def _bike_from_record(record):
parts = list(_BIKE_STRUCT.unpack(record))
parts[ID] = parts[ID].decode("utf8").rstrip("\x00")
parts[NAME] = parts[NAME].decode("utf8").rstrip("\x00")
Chapter 7. File Handling
return Bike(*parts)
def _record_from_bike(bike):
return _BIKE_STRUCT.pack(bike.identity.encode("utf8"),"utf8"),
bike.quantity, bike.price)
114 ➤
When we convert a binary record into a BikeStock.Bike we first convert the
tuple returned by unpack() into a list. This allows us to modify elements, in
this case to convert UTF-8 encoded bytes into strings with padding 0x00 bytes
stripped off. We then use the sequence unpacking operator (*) to feed the parts
to the BikeStock.Bike initializer. Packing the data is much simpler; we just
have to make sure that we encode the strings as UTF-8 bytes.
For modern desktop systems the need for application programs to use random
access binary data decreases as RAM sizes and disk speeds increase. And when
such functionality is needed, it is often easiest to use a DBM file or an SQL
database. Nonetheless, there are systems where the functionality shown here
may be useful, for example, on embedded and other resource limited systems.
This chapter showed the most widely used techniques for saving and loading
collections of data to and from files. We have seen how easy pickles are to
use, and how we can handle both compressed and uncompressed files without
knowing in advance whether compression has been used.
We saw how writing and reading binary data requires care, and saw that the
code can be quite long if we need to handle variable length strings. But we also
learned that using binary files usually results in the smallest possible file sizes
and the fastest writing and reading times. We learned too that it is important
to use a magic number to identify our file type and to use a version number to
make it practical to change the format later on.
In this chapter we saw that plain text is the easiest format for users to read and
that if the data is structured well it can be straightforward for additional tools
to be created to manipulate the data. However, parsing text data can be tricky.
We saw how to read text data both manually and using regular expressions.
XML is a very popular data interchange format and it is generally useful to be
able to at least import and export XML even when the normal format is a binary or text one. We saw how to write XML manually—including how to correctly
escape attribute values and textual data—and how to write it using an element
tree and a DOM. We also learned how to parse XML using the element tree,
DOM, and SAX parsers that Python’s standard library provides.
In the chapter’s final section we saw how to create a generic class to handle
random access binary files that hold records of a fixed size, and then how to use
the generic class in a specific context.
This chapter brings us to the end of all the fundamentals of Python programming. It is possible to stop reading right here and to write perfectly good
Python programs based on everything you have learned so far. But it would be
a shame to stop now—Python has so much more to offer, from neat techniques
that can shorten and simplify code, to some mind-bending advanced facilities
that are at least nice to know about, even if they are not often needed. In the
next chapter we will go further with procedural and object-oriented programming, and we will also get a taste of functional programming. Then, in the
following chapters we will focus more on broader programming techniques
including threading, networking, database programming, regular expressions,
and GUI (Graphical User Interface) programming.
The first exercise is to create a simpler binary record file module than the one
presented in this chapter—one whose record size is exactly the same as what
the user specifies. The second exercise is to modify the BikeStock module to
use your new binary record file module. The third exercise asks you to create
a program from scratch—the file handling is quite straightforward, but some
of the output formatting is rather challenging.
1. Make a new, simpler version of the BinaryRecordFile module—one that
does not use a state byte. For this version the record size specified by
the user is the record size actually used. New records must be added using a new append() method that simply moves the file pointer to the end
and writes the given record. The __setitem__() method should only allow
existing records to be replaced; one easy way of doing this is to use the
__seek_to_index() method. With no state byte, __getitem__() is reduced to
a mere three lines. The __delitem__() method will need to be completely
rewritten since it must move all the records up to fill the gap; this can be
done in just over half a dozen lines, but does require some thought. The
undelete() method must be removed since it is not supported, and the compact() and inplace_compact() methods must be removed because they are
no longer needed.
All told, the changes amount to fewer than 20 new or changed lines and
at least 60 deleted lines compared with the original, and not counting
doctests. A solution is provided in
2. Once you are confident that your simpler BinaryRecordFile class works,
copy the file and modify it to work with your BinaryRecordFile
Chapter 7. File Handling
class. This involves changing only a handful of lines. A solution is provided in
3. Debugging binary formats can be difficult, but a tool that can help is one
that can do a hex dump of a binary file’s contents. Create a program that
has the following console help text:
Usage: [options] file1 [file2 [... fileN]]
-h, --help
show this help message and exit
block size (8..80) [default: 16]
-d, --decimal
decimal block numbers [default: hexadecimal]
-e ENCODING, --encoding=ENCODING
encoding (ASCII..UTF-32) [default: UTF-8]
Using this program, if we have a BinaryRecordFile that is storing records
with the structure "<i10s" (little-endian, 4-byte signed integer, 10-byte
byte string), by setting the block size to match one record (15 bytes including the state byte), we can get a clear picture of what’s in the file. For example: -b15 test.dat
-------- --------------------------------00000000 02000000 00416C70 68610000 000000
00000001 01140000 00427261 766F0000 000000
00000002 02280000 00436861 726C6965 000000
00000003 023C0000 0044656C 74610000 000000
UTF-8 characters
Each byte is represented by a two-digit hexadecimal number; the spacing
between each set of four bytes (i.e., between each group of eight hexadecimal digits) is purely to improve readability. Here we can see that the second record (“Bravo”) has been deleted since its state byte is 0x01 rather
than the 0x02 used to indicate nonblank nondeleted records.
Use the optparse module to handle the command-line options. (By specifying an option’s “type” you can get optparse to handle the string-to-integer
conversion for the block size.) It can be quite tricky to get the headings
to line up correctly for any given block size and to line up the characters
correctly for the last block, so make sure you test with various block sizes
(e.g., 8, 9, 10, …, 40). Also, don’t forget that in variable length files, the last
block may be short. As the example illustrates, use periods to stand for
nonprintable characters.
The program can be written in fewer than 70 lines spread over two
functions. A solution is given in
● Further Procedural Programming
● Further Object-Oriented
● Functional-Style Programming
Advanced Programming
In this chapter we will look at a wide variety of different programming techniques and introduce many additional, often more advanced, Python syntaxes.
Some of the material in this chapter is quite challenging, but keep in mind that
the most advanced techniques are rarely needed and you can always skim the
first time to get an idea of what can be done and read more carefully when the
need arises.
The chapter’s first section digs more deeply into Python’s procedural features.
It starts by showing how to use what we already covered in a novel way, and
then returns to the theme of generators that we only touched on in Chapter 6.
The section then introduces dynamic programming—loading modules by name
at runtime and executing arbitrary code at runtime. The section returns to the
theme of local (nested) functions, but in addition covers the use of the nonlocal
keyword and recursive functions. Earlier we saw how to use Python’s predefined decorators—in this section we learn how to create our own decorators.
The section concludes with coverage of function annotations.
The second section covers all new material relating to object-oriented programming. It begins by introducing __slots__, a mechanism for minimizing the
memory used by each object. It then shows how to access attributes without using properties. The section also introduces functors (objects that can be called
like functions), and context managers—these are used in conjunction with the
with keyword, and in many cases (e.g., file handling) they can be used to replace
try … except … finally constructs with simpler try … except constructs. The
section also shows how to create custom context managers, and introduces additional advanced object-oriented features, including class decorators, abstract
base classes, multiple inheritance, and metaclasses.
The third section introduces some fundamental concepts of functional programming, and introduces some useful functions from the functools, itertools,
Chapter 8. Advanced Programming Techniques
and operator modules. This section also shows how to use partial function application to simplify code, and how to create and use coroutines.
All the previous chapters put together have provided us with the “standard
Python toolbox”. This chapter takes everything that we have already covered
and turns it into the “deluxe Python toolbox”, with all the original tools (techniques and syntaxes), plus many new ones that can make our programming
easier, shorter, and more effective. Some of the tools can have interchangeable
uses, for example, some jobs can be done using either a class decorator or a
metaclass, whereas others, such as descriptors, can be used in multiple ways to
achieve different effects. Some of the tools covered here, for example, context
managers, we will use all the time, and others will remain ready at hand for
those particular situations for which they are the perfect solution.
Further Procedural Programming
Most of this section deals with additional facilities relating to procedural
programming and functions, but the very first subsection is different in that it
presents a useful programming technique based on what we already covered
without introducing any new syntax.
Branching Using Dictionaries
As we noted earlier, functions are objects like everything else in Python, and
a function’s name is an object reference that refers to the function. If we write
a function’s name without parentheses, Python knows we mean the object
reference, and we can pass such object references around just like any others.
We can use this fact to replace if statements that have lots of elif clauses with
a single function call.
In Chapter 12 we will review an interactive console program called,
that has the following menu:
The program has a function that gets the user’s choice and which will return
only a valid choice, in this case one of “a”, “e”, “l”, “r”, “i”, “x”, and “q”. Here are
two equivalent code snippets for calling the relevant function based on the
user’s choice:
if action == "a":
elif action == "e":
Further Procedural Programming
elif action == "l":
elif action == "r":
elif action == "i":
elif action == "x":
elif action == "q":
functions = dict(a=add_dvd, e=edit_dvd,
l=list_dvds, r=remove_dvd,
i=import_, x=export, q=quit)
The choice is held as a one-character string in the action variable, and the
database to be used is held in the db variable. The import_() function has a
trailing underscore to keep it distinct from the built-in import statement.
In the right-hand code snippet we create a dictionary whose keys are the valid
menu choices, and whose values are function references. In the second statement we retrieve the function reference corresponding to the given action and
call the function referred to using the call operator, (), and in this example,
passing the db argument. Not only is the code on the right-hand side much
shorter than the code on the left, but also it can scale (have far more dictionary items) without affecting its performance, unlike the left-hand code whose
speed depends on how many elifs must be tested to find the appropriate function to call.
The program from the preceding chapter uses this
technique in its import_() method, as this extract from the method shows:
call = {(".aix", "dom"): self.import_xml_dom,
(".aix", "etree"): self.import_xml_etree,
(".aix", "sax"): self.import_xml_sax,
(".ait", "manual"): self.import_text_manual,
(".ait", "regex"): self.import_text_regex,
(".aib", None): self.import_binary,
(".aip", None): self.import_pickle}
result = call[extension, reader](filename)
The complete method is 13 lines long; the extension parameter is computed in
the method, and the reader is passed in. The dictionary keys are 2-tuples, and
the values are methods. If we had used if statements, the code would be 22
lines long, and would not scale as well.
Generator Expressions and Functions
Generator functions
279 ➤
Back in Chapter 6 we introduced generator functions and methods. It is
also possible to create generator expressions. These are syntactically almost
Chapter 8. Advanced Programming Techniques
identical to list comprehensions, the difference being that they are enclosed in
parentheses rather than brackets. Here are their syntaxes:
(expression for item in iterable)
(expression for item in iterable if condition)
In the preceding chapter we created some iterator methods using yield
expressions. Here are two equivalent code snippets that show how a simple for
… in loop containing a yield expression can be coded as a generator:
def items_in_key_order(d):
for key in sorted(d):
yield key, d[key]
def items_in_key_order(d):
return ((key, d[key])
for key in sorted(d))
Both functions return a generator that produces a list of key–value items
for the given dictionary. If we need all the items in one go we can pass the
generator returned by the functions to list() or tuple(); otherwise, we can
iterate over the generator to retrieve items as we need them.
Generators provide a means of performing lazy evaluation, which means that
they compute only the values that are actually needed. This can be more efficient than, say, computing a very large list in one go. Some generators produce
as many values as we ask for—without any upper limit. For example:
def quarters(next_quarter=0.0):
while True:
yield next_quarter
next_quarter += 0.25
This function will return 0.0, 0.25, 0.5, and so on, forever. Here is how we could
use the generator:
result = []
for x in quarters():
if x >= 1.0:
The break statement is essential—without it the for … in loop will never finish.
At the end the result list is [0.0, 0.25, 0.5, 0.75, 1.0].
Every time we call quarters() we get back a generator that starts at 0.0 and
increments by 0.25; but what if we want to reset the generator’s current
value? It is possible to pass a value into a generator, as this new version of the
generator function shows:
def quarters(next_quarter=0.0):
while True:
Further Procedural Programming
received = (yield next_quarter)
if received is None:
next_quarter += 0.25
next_quarter = received
The yield expression returns each value to the caller in turn. In addition, if
the caller calls the generator’s send() method, the value sent is received in the
generator function as the result of the yield expression. Here is how we can
use the new generator function:
result = []
generator = quarters()
while len(result) < 5:
x = next(generator)
if abs(x - 0.5) < sys.float_info.epsilon:
x = generator.send(1.0)
We create a variable to refer to the generator and call the built-in next() function which retrieves the next item from the generator it is given. (The same
effect can be achieved by calling the generator’s __next__() special method, in
this case, x = generator.__next__().) If the value is equal to 0.5 we send the value
1.0 into the generator (which immediately yields this value back). This time the
result list is [0.0, 0.25, 1.0, 1.25, 1.5].
In the next subsection we will review the program which processes files given on the command line. Unfortunately, the Windows shell program (cmd.exe) does not provide wildcard expansion (also called file globbing), so
if a program is run on Windows with the argument *.*, the literal text “*.*” will
go into the sys.argv list instead of all the files in the current directory. We solve
this problem by creating two different get_files() functions, one for Windows
and the other for Unix, both of which use generators. Here’s the code:
if sys.platform.startswith("win"):
def get_files(names):
for name in names:
if os.path.isfile(name):
yield name
for file in glob.iglob(name):
if not os.path.isfile(file):
yield file
def get_files(names):
return (file for file in names if os.path.isfile(file))
Chapter 8. Advanced Programming Techniques
In either case the function is expected to be called with a list of filenames, for
example, sys.argv[1:], as its argument.
On Windows the function iterates over all the names listed. For each filename,
the function yields the name, but for nonfiles (usually directories), the glob
module’s glob.iglob() function is used to return an iterator to the names of the
files that the name represents after wildcard expansion. For an ordinary name
like autoexec.bat an iterator that produces one item (the name) is returned,
and for a name that uses wildcards like *.txt an iterator that produces all the
matching files (in this case those with extension .txt) is returned. (There is
also a glob.glob() function that returns a list rather than an iterator.)
On Unix the shell does wildcard expansion for us, so we just need to return a
generator for all the files whose names we have been given.★
Generator functions can also be used as coroutines, if we structure them
correctly. Coroutines are functions that can be suspended in mid-execution
(at the yield expression), waiting for the yield to provide a result to work on,
and once received they continue processing. As we will see in the coroutines
subsection later in this chapter (➤ 399), coroutines can be used to distribute
work and to create processing pipelines.
Dynamic Code Execution and Dynamic Imports
There are some occasions when it is easier to write a piece of code that generates the code we need than to write the needed code directly. And in some
contexts it is useful to let users enter code (e.g., functions in a spreadsheet),
and to let Python execute the entered code for us rather than to write a parser
and handle it ourselves—although executing arbitrary code like this is a potential security risk, of course. Another use case for dynamic code execution
is to provide plug-ins to extend a program’s functionality. Using plug-ins has
the disadvantage that all the necessary functionality is not built into the program (which can make the program more difficult to deploy and runs the risk
of plug-ins getting lost), but has the advantages that plug-ins can be upgraded
individually and can be provided separately, perhaps to provide enhancements
that were not originally envisaged.
Dynamic Code Execution
The easiest way to execute an expression is to use the built-in eval() function
we first saw in Chapter 6. For example:
x = eval("(2 ** 31) - 1")
# x == 2147483647
The glob.glob() functions are not as powerful as, say, the Unix bash shell, since although they
support the *, ?, and [] syntaxes, they don’t support the {} syntax.
Further Procedural Programming
This is fine for user-entered expressions, but what if we need to create a
function dynamically? For that we can use the built-in exec() function. For
example, the user might give us a formula such as 4π r2 and the name “area of
sphere”, which they want turned into a function. Assuming that we replace π
with math.pi, the function they want can be created like this:
import math
code = '''
def area_of_sphere(r):
return 4 * math.pi * r ** 2
context = {}
context["math"] = math
exec(code, context)
We must use proper indentation—after all, the quoted code is standard Python.
(Although in this case we could have written it all on a single line because the
suite is just one line.)
If exec() is called with some code as its only argument there is no way to
access any functions or variables that are created as a result of the code being
executed. Furthermore, exec() cannot access any imported modules or any of
the variables, functions, or other objects that are in scope at the point of the
call. Both of these problems can be solved by passing a dictionary as the second
argument. The dictionary provides a place where object references can be kept
for accessing after the exec() call has finished. For example, the use of the
context dictionary means that after the exec() call, the dictionary has an object
reference to the area_of_sphere() function that was created by exec(). In this
example we needed exec() to be able to access the math module, so we inserted
an item into the context dictionary whose key is the module’s name and whose
value is an object reference to the corresponding module object. This ensures
that inside the exec() call, math.pi is accessible.
In some cases it is convenient to provide the entire global context to exec().
This can be done by passing the dictionary returned by the globals() function.
One disadvantage of this approach is that any objects created in the exec() call
would be added to the global dictionary. A solution is to copy the global context
into a dictionary, for example, context = globals().copy(). This still gives exec()
access to imported modules and the variables and other objects that are in
scope, and because we have copied, any changes to the context made inside the
exec() call are kept in the context dictionary and are not propagated to the global environment. (It would appear to be more secure to use copy.deepcopy(), but
if security is a concern it is best to avoid exec() altogether.) We can also pass
the local context, for example, by passing locals() as a third argument—this
makes objects in the local scope accessible to the code executed by exec().
Chapter 8. Advanced Programming Techniques
After the exec() call the context dictionary contains a key called "area_of_
sphere" whose value is the area_of_sphere() function. Here is how we can
access and call the function:
area_of_sphere = context["area_of_sphere"]
area = area_of_sphere(5)
# area == 314.15926535897933
The area_of_sphere object is an object reference to the function we have dynamically created and can be used just like any other function. And although we
created only a single function in the exec() call, unlike eval(), which can operate on only a single expression, exec() can handle as many Python statements
as we like, including entire modules, as we will see in the next subsubsection.
Dynamically Importing Modules
Python provides three straightforward mechanisms that can be used to create
plug-ins, all of which involve importing modules by name at runtime. And
once we have dynamically imported additional modules, we can use Python’s
introspection functions to check the availability of the functionality we want,
and to access it as required.
In this subsubsection we will review the program. This
program reads the first 1 000 bytes of each file given on the command line and
for each one outputs the file’s type (or the text “Unknown”), and the filename.
Here is an example command line and an extract from its output:
C:\Python31\python.exe c:\windows\*.*
Windows Executable..c:\windows\winhelp.exe
Windows Executable..c:\windows\winhlp32.exe
Windows BMP Image...c:\windows\winnt.bmp
The program tries to load in any module that is in the same directory as the
program and whose name contains the text “magic”.Such modules are expected
to provide a single public function, get_file_type(). Two very simple example
modules, and, that each have
a get_file_type() function are provided with the book’s examples.
We will review the program’s main() function in two parts.
def main():
modules = load_modules()
get_file_type_functions = []
for module in modules:
Further Procedural Programming
get_file_type = get_function(module, "get_file_type")
if get_file_type is not None:
In a moment, we will look at three different implementations of the
load_modules() function which returns a (possibly empty) list of module objects,
and we will look at the get_function() function further on. For each module
found we try to retrieve a get_file_type() function, and add any we get to a list
of such functions.
for file in get_files(sys.argv[1:]):
fh = None
fh = open(file, "rb")
magic =
for get_file_type in get_file_type_functions:
filetype = get_file_type(magic,
if filetype is not None:
print("{0:.<20}{1}".format(filetype, file))
print("{0:.<20}{1}".format("Unknown", file))
except EnvironmentError as err:
if fh is not None:
This loop iterates over every file listed on the command line and for each one
reads its first 1 000 bytes. It then tries each get_file_type() function in turn
to see whether it can determine the current file’s type. If the file type is determined, the details are printed and the inner loop is broken out of, with processing continuing with the next file. If no function can determine the file type—or
if no get_file_type() functions were found—an “Unknown” line is printed.
We will now review three different (but equivalent) ways of dynamically
importing modules, starting with the longest and most difficult approach, since
it shows every step explicitly:
def load_modules():
modules = []
for name in os.listdir(os.path.dirname(__file__) or "."):
if name.endswith(".py") and "magic" in name.lower():
filename = name
name = os.path.splitext(name)[0]
if name.isidentifier() and name not in sys.modules:
Chapter 8. Advanced Programming Techniques
fh = None
fh = open(filename, "r", encoding="utf8")
code =
module = type(sys)(name)
sys.modules[name] = module
exec(code, module.__dict__)
except (EnvironmentError, SyntaxError) as err:
sys.modules.pop(name, None)
if fh is not None:
return modules
We begin by iterating over all the files in the program’s directory. If this is the
current directory, os.path.dirname(__file__) will return an empty string which
would cause os.listdir() to raise an exception, so we pass "." if necessary.
For each candidate file (ends with .py and contains the text “magic”), we get
the module name by chopping off the file extension. If the name is a valid
identifier it is a viable module name, and if it isn’t already in the global list of
modules maintained in the sys.modules dictionary we can try to import it.
We read the text of the file into the code string. The next line, module =
type(sys)(name), is quite subtle. When we call type() it returns the type object
of the object it is given. So if we called type(1) we would get int back. If we
print the type object we just get something human readable like “int”, but if
we call the type object as a function, we get an object of that type back. For
example, we can get the integer 5 in variable x by writing x = 5, or x = int(5),
or x = type(0)(5), or int_type = type(0); x = int_type(5). In this case we’ve used
type(sys) and sys is a module, so we get back the module type object (essentially
the same as a class object), and can use it to create a new module with the given name. Just as with the int example where it didn’t matter what integer we
used to get the int type object, it doesn’t matter what module we use (as long as
it is one that exists, that is, has been imported) to get the module type object.
Once we have a new (empty) module, we add it to the global list of modules to
prevent the module from being accidentally reimported. This is done before
calling exec() to more closely mimic the behavior of the import statement.
Then we call exec() to execute the code we have read—and we use the module’s
dictionary as the code’s context. At the end we add the module to the list of
modules we will pass back. And if a problem arises, we delete the module from
the global modules dictionary if it has been added—it will not have been added
to the list of modules if an error occurred. Notice that exec() can handle any
Further Procedural Programming
Table 8.1 Dynamic Programming and Introspection Functions
Imports a module by name; see text
Returns the code object that results from compiling the
source text; file should be the filename, or "<string>";
mode must be “single”, “eval”, or “exec”
Deletes the attribute called name from object obj
Returns the list of names in the local scope, or if obj is
given then obj’s names (e.g., its attributes and methods)
Returns the result of evaluating the single expression in
source; if supplied, globals is the global context and locals
is the local context (as dictionaries)
Evaluates object obj, which can be a string or a code object
from compile(), and returns None; if supplied, globals is
the global context and locals is the local context
name, val)
Returns the value of the attribute called name from object
obj, or val if given and there is no such attribute
Returns a dictionary of the current global context
Returns True if object obj has an attribute called name
Returns a dictionary of the current local context
name, val)
Sets the attribute called name to the value val for the object
obj, creating the attribute if necessary
Returns object obj’s type object
Returns object obj’s context as a dictionary; or the local
context if obj is not given
amount of code (whereas eval() evaluates a single expression—see Table 8.1),
and raises a SyntaxError exception if there’s a syntax error.
Here’s the second way to dynamically load a module at runtime—the code
shown here replaces the first approach’s try … except block:
exec("import " + name)
except SyntaxError as err:
Chapter 8. Advanced Programming Techniques
One theoretical problem with this approach is that it is potentially insecure.
The name variable could begin with sys; and be followed by some destructive code.
And here is the third approach, again just showing the replacement for the first
approach’s try … except block:
module = __import__(name)
except (ImportError, SyntaxError) as err:
This is the easiest way to dynamically import modules and is slightly safer
than using exec(), although like any dynamic import, it is by no means secure
because we don’t know what is being executed when the module is imported.
None of the techniques shown here handles packages or modules in different
paths, but it is not difficult to extend the code to accommodate these—although
it is worth reading the online documentation, especially for __import__(), if
more sophistication is required.
Having imported the module we need to be able to access the functionality it
provides. This can be achieved using Python’s built-in introspection functions,
getattr() and hasattr(). Here’s how we have used them to implement the
get_function() function:
def get_function(module, function_name):
function = get_function.cache.get((module, function_name), None)
if function is None:
function = getattr(module, function_name)
if not hasattr(function, "__call__"):
raise AttributeError()
get_function.cache[module, function_name] = function
except AttributeError:
function = None
return function
get_function.cache = {}
Ignoring the cache-related code for a moment, what the function does is call
getattr() on the module object with the name of the function we want. If
there is no such attribute an AttributeError exception is raised, but if there
is such an attribute we use hasattr() to check that the attribute itself has
the __call__ attribute—something that all callables (functions and methods)
have. (Further on we will see a nicer way of checking whether an attribute is
➤ 392
Chapter 8. Advanced Programming Techniques
One common use case for local functions is when we want to use recursion. In
these cases, the enclosing function is called, sets things up, and then makes
the first call to a local recursive function. Recursive functions (or methods) are
ones that call themselves. Structurally, all directly recursive functions can be
seen as having two cases: the base case and the recursive case. The base case is
used to stop the recursion.
Recursive functions can be computationally expensive because for every recursive call another stack frame is used; however, some algorithms are most
naturally expressed using recursion. Most Python implementations have a
fixed limit to how many recursive calls can be made. The limit is returned by
sys.getrecursionlimit() and can be changed by sys.setrecursionlimit(), although increasing the limit is most often a sign that the algorithm being used
is inappropriate or that the implementation has a bug.
The classic example of a recursive function is one that is used to calculate
factorials.★ For example, factorial(5) will calculate 5! and return 120, that is,
1 × 2 × 3 × 4 × 5:
def factorial(x):
if x <= 1:
return 1
return x * factorial(x - 1)
This is not an efficient solution, but it does show the two fundamental features
of recursive functions. If the given number, x, is 1 or less, 1 is returned and
no recursion occurs—this is the base case. But if x is greater than 1 the value
returned is x * factorial(x - 1), and this is the recursive case because here the
factorial function calls itself. The function is guaranteed to terminate because
if the initial x is less than or equal to 1 the base case will be used and the
function will finish immediately, and if x is greater than 1, each recursive call
will be on a number one less than before and so will eventually be 1.
To see both local functions and recursive functions in a meaningful context we
will study the indented_list_sort() function from module file
This function takes a list of strings that use indentation to create a hierarchy,
and a string that holds one level of indent, and returns a list with the same
strings but where all the strings are sorted in case-insensitive alphabetical
order, with indented items sorted under their parent item, recursively, as the
before and after lists shown in Figure 8.1 illustrate.
Given the before list, the after list is produced by this call: after = IndentedList.indented_list_sort(before). The default indent value is four spaces, the
same as the indent used in the before list, so we did not need to set it explicitly.
Python’s math module provides a much more efficient math.factorial() function.
Further Procedural Programming
before = ["Nonmetals",
"Inner Transitionals",
"Alkali Metals",
after = ["Alkali Metals",
"Inner Transitionals",
Figure 8.1 Before and after sorting an indented list
We will begin by looking at the indented_list_sort() function as a whole, and
then we will look at its two local functions.
def indented_list_sort(indented_list, indent="
KEY, ITEM, CHILDREN = range(3)
def add_entry(level, key, item, children):
def update_indented_list(entry):
entries = []
for item in indented_list:
level = 0
i = 0
while item.startswith(indent, i):
i += len(indent)
level += 1
key = item.strip().lower()
add_entry(level, key, item, entries)
indented_list = []
for entry in sorted(entries):
return indented_list
Chapter 8. Advanced Programming Techniques
The code begins by creating three constants that are used to provide names for
index positions used by the local functions. Then we define the two local functions which we will review in a moment. The sorting algorithm works in two
stages. In the first stage we create a list of entries, each a 3-tuple consisting of
a “key” that will be used for sorting, the original string, and a list of the string’s
child entries. The key is just a lowercased copy of the string with whitespace
stripped from both ends. The level is the indentation level, 0 for top-level items,
1 for children of top-level items, and so on. In the second stage we create a new
indented list and add each string from the sorted entries list, and each string’s
child strings, and so on, to produce a sorted indented list.
def add_entry(level, key, item, children):
if level == 0:
children.append((key, item, []))
add_entry(level - 1, key, item, children[-1][CHILDREN])
This function is called for each string in the list. The children argument is the
list to which new entries must be added. When called from the outer function
(indented_list_sort()), this is the entries list. This has the effect of turning a
list of strings into a list of entries, each of which has a top-level (unindented)
string and a (possibly empty) list of child entries.
If the level is 0 (top-level), we add a new 3-tuple to the entries list. This holds
the key (for sorting), the original item (which will go into the resultant sorted
list), and an empty children list. This is the base case since no recursion takes
place. If the level is greater than 0, the item is a child (or descendant) of the
last item in the children list. In this case we recursively call add_entry() again,
reducing the level by 1 and passing the children list’s last item’s children list as
the list to add to. If the level is 2 or more, more recursive calls will take place,
until eventually the level is 0 and the children list is the right one for the entry
to be added to.
For example, when the “Inner Transitionals” string is reached, the outer function calls add_entry() with a level of 0, a key of “inner transitionals”, an item of
“Inner Transitionals”, and the entries list as the children list. Since the level is
0, a new item will be appended to the children list (entries), with the key, item,
and an empty children list. The next string is “ Lanthanides”—this is indented, so it is a child of the “Inner Transitionals” string. The add_entry() call this
time has a level of 1, a key of “lanthanides”, an item of “ Lanthanides”, and
the entries list as the children list. Since the level is 1, the add_entry() function
calls itself recursively, this time with level 0 (1 - 1), the same key and item, but
with the children list being the children list of the last item, that is, the “Inner
Transitionals” item’s children list.
Here is what the entries list looks like once all the strings have been added, but
before the sorting has been done:
Further Procedural Programming
[('hydrogen', '
Hydrogen', []),
('carbon', '
Carbon', []),
('nitrogen', '
Nitrogen', []),
('oxygen', '
Oxygen', [])]),
('inner transitionals',
'Inner Transitionals',
[('cerium', '
Cerium', []),
('europium', '
Europium', [])]),
[('uranium', '
Uranium', []),
('curium', '
Curium', []),
('plutonium', '
Plutonium', [])])]),
('alkali metals',
'Alkali Metals',
[('lithium', '
Lithium', []),
('sodium', '
Sodium', []),
('potassium', '
Potassium', [])])]
The output was produced using the pprint (“pretty print”) module’s pprint.
pprint() function. Notice that the entries list has only three items (all of which
are 3-tuples), and that each 3-tuple’s last element is a list of child 3-tuples (or
is an empty list).
The add_entry() function is both a local function and a recursive function. Like
all recursive functions, it has a base case (in this function, when the level is 0)
that ends the recursion, and a recursive case.
The function could be written in a slightly different way:
def add_entry(key, item, children):
nonlocal level
if level == 0:
children.append((key, item, []))
level -= 1
add_entry(key, item, children[-1][CHILDREN])
Here, instead of passing level as a parameter, we use a nonlocal statement to
access a variable in an outer enclosing scope. If we did not change level inside
the function we would not need the nonlocal statement—in such a situation,
Python would not find it in the local (inner function) scope, and would look
at the enclosing scope and find it there. But in this version of add_entry() we
Chapter 8. Advanced Programming Techniques
need to change level’s value, and just as we need to tell Python that we want
to change global variables using the global statement (to prevent a new local
variable from being created rather than the global variable updated), the same
applies to variables that we want to change but which belong to an outer scope.
Although it is often best to avoid using global altogether, it is also best to use
nonlocal with care.
def update_indented_list(entry):
for subentry in sorted(entry[CHILDREN]):
In the algorithm’s first stage we build up a list of entries, each a (key, item,
children) 3-tuple, in the same order as they are in the original list. In the
algorithm’s second stage we begin with a new empty indented list and iterate
over the sorted entries, calling update_indented_list() for each one to build up
the new indented list. The update_indented_list() function is recursive. For
each top-level entry it adds an item to the indented_list, and then calls itself
for each of the item’s child entries. Each child is added to the indented_list,
and then the function calls itself for each child’s children—and so on. The base
case (when the recursion stops) is when an item, or child, or child of a child,
and so on has no children of its own.
Python looks for indented_list in the local (inner function) scope and doesn’t
find it, so it then looks in the enclosing scope and finds it there. But notice that
inside the function we append items to the indented_list even though we have
not used nonlocal. This works because nonlocal (and global) are concerned with
object references, not with the objects they refer to. In the second version of
add_entry() we had to use nonlocal for level because the += operator applied to
a number rebinds the object reference to a new object—what really happens is
level = level + 1, so level is set to refer to a new integer object. But when we call
list.append() on the indented_list, it modifies the list itself and no rebinding
takes place, and therefore nonlocal is not necessary. (For the same reason, if
we have a dictionary, list, or other global collection, we can add or remove items
from it without using a global statement.)
Function and Method Decorators
A decorator is a function that takes a function or method as its sole argument
and returns a new function or method that incorporates the decorated function
or method with some additional functionality added. We have already made
use of some predefined decorators, for example, @property and @classmethod. In
this subsection we will learn how to create our own function decorators, and
later in this chapter we will see how to create class decorators.
➤ 378
Further Procedural Programming
For our first decorator example, let us suppose that we have many functions
that perform calculations, and that some of these must always produce a positive result. We could add an assertion to each of these, but using a decorator is
easier and clearer. Here’s a function decorated with the @positive_result decorator that we will create in a moment:
def discriminant(a, b, c):
return (b ** 2) - (4 * a * c)
Thanks to the decorator, if the result is ever less than 0, an AssertionError exception will be raised and the program will terminate. And of course, we can
use the decorator on as many functions as we like. Here’s the decorator’s implementation:
def positive_result(function):
def wrapper(*args, **kwargs):
result = function(*args, **kwargs)
assert result >= 0, function.__name__ + "() result isn't >= 0"
return result
wrapper.__name__ = function.__name__
wrapper.__doc__ = function.__doc__
return wrapper
Decorators define a new local function that calls the original function. Here,
the local function is wrapper(); it calls the original function and stores the
result, and it uses an assertion to guarantee that the result is positive (or that
the program will terminate). The wrapper finishes by returning the result
computed by the wrapped function. After creating the wrapper, we set its name
and docstring to those of the original function. This helps with introspection,
since we want error messages to mention the name of the original function, not
the wrapper. Finally, we return the wrapper function—it is this function that
will be used in place of the original.
def positive_result(function):
def wrapper(*args, **kwargs):
result = function(*args, **kwargs)
assert result >= 0, function.__name__ + "() result isn't >= 0"
return result
return wrapper
Here is a slightly cleaner version of the @positive_result decorator. The wrapper itself is wrapped using the functools module’s @functools.wraps decorator,
which ensures that the wrapper() function has the name and docstring of the
original function.
Chapter 8. Advanced Programming Techniques
In some cases it would be useful to be able to parameterize a decorator, but at
first sight this does not seem possible since a decorator takes just one argument, a function or method. But there is a neat solution to this. We can call a
function with the parameters we want and that returns a decorator which can
then decorate the function that follows it. For example:
@bounded(0, 100)
def percent(amount, total):
return (amount / total) * 100
Here, the bounded() function is called with two arguments, and returns a decorator that is used to decorate the percent() function. The purpose of the decorator in this case is to guarantee that the number returned is always in the range
0 to 100 inclusive. Here’s the implementation of the bounded() function:
def bounded(minimum, maximum):
def decorator(function):
def wrapper(*args, **kwargs):
result = function(*args, **kwargs)
if result < minimum:
return minimum
elif result > maximum:
return maximum
return result
return wrapper
return decorator
The function creates a decorator function, that itself creates a wrapper function. The wrapper performs the calculation and returns a result that is within
the bounded range. The decorator() function returns the wrapper() function,
and the bounded() function returns the decorator.
One further point to note is that each time a wrapper is created inside the
bounded() function, the particular wrapper uses the minimum and maximum
values that were passed to bounded().
The last decorator we will create in this subsection is a bit more complex. It is a
logging function that records the name, arguments, and result of any function
it is used to decorate. For example:
def discounted_price(price, percentage, make_integer=False):
result = price * ((100 - percentage) / 100)
if not (0 < result <= price):
raise ValueError("invalid price")
return result if not make_integer else int(round(result))
Further Procedural Programming
If Python is run in debug mode (the normal mode), every time the discounted_price() function is called a log message will be added to the file logged.log
in the machine’s local temporary directory, as this log file extract illustrates:
10) -> 90.0
5) -> 199.5
5, make_integer=True) -> 200
14, True) -> 181
-8) <type 'ValueError'>: invalid price
If Python is run in optimized mode (using the -O command-line option or if
the PYTHONOPTIMIZE environment variable is set to -O), then no logging will take
place. Here’s the code for setting up logging and for the decorator:
if __debug__:
logger = logging.getLogger("Logger")
handler = logging.FileHandler(os.path.join(
tempfile.gettempdir(), "logged.log"))
def logged(function):
def wrapper(*args, **kwargs):
log = "called: " + function.__name__ + "("
log += ", ".join(["{0!r}".format(a) for a in args] +
["{0!s}={1!r}".format(k, v)
for k, v in kwargs.items()])
result = exception = None
result = function(*args, **kwargs)
return result
except Exception as err:
exception = err
log += ((") -> " + str(result)) if exception is None
else ") {0}: {1}".format(type(exception),
if exception is not None:
raise exception
return wrapper
def logged(function):
return function
Chapter 8. Advanced Programming Techniques
In debug mode the global variable __debug__ is True. If this is the case we set up
logging using the logging module, and then create the @logged decorator. The
logging module is very powerful and flexible—it can log to files, rotated files,
emails, network connections, HTTP servers, and more. Here we’ve used only
the most basic facilities by creating a logging object, setting its logging level
(several levels are supported), and choosing to use a file for the output.
134 ➤
The wrapper’s code begins by setting up the log string with the function’s name
and arguments. We then try calling the function and storing its result. If any
exception occurs we store it. In all cases the finally block is executed, and
there we add the return value (or exception) to the log string and write to the
log. If no exception occurred, the result is returned; otherwise, we reraise the
exception to correctly mimic the original function’s behavior.
If Python is running in optimized mode, __debug__ is False; in this case we
define the logged() function to simply return the function it is given, so apart
from the tiny overhead of this indirection when the function is first created,
there is no runtime overhead at all.
Note that the standard library’s trace and cProfile modules can run and analyse programs and modules to produce various tracing and profiling reports.
Both use introspection, so unlike the @logged decorator we have used here, neither trace nor cProfile requires any source code changes.
Function Annotations
Functions and methods can be defined with annotations—expressions that can
be used in a function’s signature. Here’s the general syntax:
def functionName(par1 : exp1, par2 : exp2, ..., parN : expN) -> rexp:
Every colon expression part (: expX) is an optional annotation, and so is the
arrow return expression part (-> rexp). The last (or only) positional parameter
(if present) can be of the form *args, with or without an annotation; similarly,
the last (or only) keyword parameter (if present) can be of the form **kwargs,
again with or without an annotation.
If annotations are present they are added to the function’s __annotations__ dictionary; if they are not present this dictionary is empty. The dictionary’s keys
are the parameter names, and the values are the corresponding expressions.
The syntax allows us to annotate all, some, or none of the parameters and to
annotate the return value or not. Annotations have no special significance to
Python. The only thing that Python does in the face of annotations is to put
them in the __annotations__ dictionary; any other action is up to us. Here is an
example of an annotated function that is in the Util module:
Further Procedural Programming
def is_unicode_punctuation(s : str) -> bool:
for c in s:
if unicodedata.category(c)[0] != "P":
return False
return True
Every Unicode character belongs to a particular category and each category is
identified by a two-character identifier. All the categories that begin with P are
punctuation characters.
Here we have used Python data types as the annotation expressions. But they
have no particular meaning for Python, as these calls should make clear:
Util.is_unicode_punctuation(s="[email protected]#?")
Util.is_unicode_punctuation(("!", "@"))
# returns: False
# returns: True
# returns: True
The first call uses a positional argument and the second call a keyword argument, just to show that both kinds work as expected. The last call passes a
tuple rather than a string, and this is accepted since Python does nothing more
than record the annotations in the __annotations__ dictionary.
If we want to give meaning to annotations, for example, to provide type checking, one approach is to decorate the functions we want the meaning to apply to
with a suitable decorator. Here is a very basic type-checking decorator:
def strictly_typed(function):
annotations = function.__annotations__
arg_spec = inspect.getfullargspec(function)
assert "return" in annotations, "missing type for return value"
for arg in arg_spec.args + arg_spec.kwonlyargs:
assert arg in annotations, ("missing type for parameter '" +
arg + "'")
def wrapper(*args, **kwargs):
for name, arg in (list(zip(arg_spec.args, args)) +
assert isinstance(arg, annotations[name]), (
"expected argument '{0}' of {1} got {2}".format(
name, annotations[name], type(arg)))
result = function(*args, **kwargs)
assert isinstance(result, annotations["return"]), (
"expected return of {0} got {1}".format(
annotations["return"], type(result)))
return result
return wrapper
Chapter 8. Advanced Programming Techniques
This decorator requires that every argument and the return value must be
annotated with the expected type. It checks that the function’s arguments and
return type are all annotated with their types when the function it is passed is
created, and at runtime it checks that the types of the actual arguments match
those expected.
The inspect module provides powerful introspection services for objects. Here,
we have made use of only a small part of the argument specification object
it returns, to get the names of each positional and keyword argument—in
the correct order in the case of the positional arguments. These names are
then used in conjunction with the annotations dictionary to ensure that every
parameter and the return value are annotated.
The wrapper function created inside the decorator begins by iterating over
every name–argument pair of the given positional and keyword arguments.
Since zip() returns an iterator and dictionary.items() returns a dictionary
view we cannot concatenate them directly, so first we convert them both to lists.
If any actual argument has a different type from its corresponding annotation
the assertion will fail; otherwise, the actual function is called and the type of
the value returned is checked, and if it is of the right type, it is returned. At the
end of the strictly_typed() function, we return the wrapped function as usual.
Notice that the checking is done only in debug mode (which is Python’s default
mode—controlled by the -O command-line option and the PYTHONOPTIMIZE environment variable).
If we decorate the is_unicode_punctuation() function with the @strictly_typed
decorator, and try the same examples as before using the decorated version, the
annotations are acted upon:
is_unicode_punctuation(s="[email protected]#?")
is_unicode_punctuation(("!", "@"))
# returns: False
# returns: True
# raises AssertionError
Now the argument types are checked, so in the last case an AssertionError is
raised because a tuple is not a string or a subclass of str.
Now we will look at a completely different use of annotations. Here’s a small
function that has the same functionality as the built-in range() function, except
that it always returns floats:
def range_of_floats(*args) -> "author=Reginald Perrin":
return (float(x) for x in range(*args))
No use is made of the annotation by the function itself, but it is easy to envisage
a tool that imported all of a project’s modules and produced a list of function
names and author names, extracting each function’s name from its __name__
attribute, and the author names from the value of the __annotations__ dictionary’s "return" item.
Further Procedural Programming
Annotations are a very new feature of Python, and because Python does not
impose any predefined meaning on them, the uses they can be put to are limited only by our imagination. Further ideas for possible uses, and some useful
links, are available from PEP 3107 “Function Annotations”,
Further Object-Oriented Programming
In this section we will look more deeply into Python’s support for object
orientation, learning many techniques that can reduce the amount of code we
must write, and that expand the power and capabilities of the programming
features that are available to us. But we will begin with one very small and
simple new feature. Here is the start of the definition of a Point class that has
exactly the same behavior as the versions we created in Chapter 6:
class Point:
__slots__ = ("x", "y")
def __init__(self, x=0, y=0):
self.x = x
self.y = y
349 ➤
When a class is created without the use of __slots__, behind the scenes Python
creates a private dictionary called __dict__ for each instance, and this dictionary holds the instance’s data attributes. This is why we can add or remove attributes from objects. (For example, we added a cache attribute to the
get_function() function earlier in this chapter.)
If we only need objects where we access the original attributes and don’t need
to add or remove attributes, we can create classes that don’t have a __dict__.
This is achieved simply by defining a class attribute called __slots__ whose
value is a tuple of attribute names. Each object of such a class will have
attributes of the specified names and no __dict__; no attributes can be added or
removed from such classes. These objects consume less memory and are faster
than conventional objects, although this is unlikely to make much difference
unless large numbers of objects are created. If we inherit from a class that uses
__slots__ we must declare slots in our subclass, even if empty, such as __slots__
= (); or the memory and speed savings will be lost.
Controlling Attribute Access
It is sometimes convenient to have a class where attribute values are computed
on the fly rather than stored. Here’s the complete implementation of such
a class:
Chapter 8. Advanced Programming Techniques
class Ord:
def __getattr__(self, char):
return ord(char)
With the Ord class available, we can create an instance, ord = Ord(), and then
have an alternative to the built-in ord() function that works for any character
that is a valid identifier. For example, ord.a returns 97, ord.Z returns 90, and
ord.å returns 229. (But ord.! and similar are syntax errors.)
Note that if we typed the Ord class into IDLE it would not work if we then typed
ord = Ord(). This is because the instance has the same name as the built-in ord()
function that the Ord class uses, so the ord() call would actually become a call
to the ord instance and result in a TypeError exception. The problem would not
arise if we imported a module containing the Ord class because the interactively
created ord object and the built-in ord() function used by the Ord class would be
in two separate modules, so one would not displace the other. If we really need
to create a class interactively and to reuse the name of a built-in we can do so by
ensuring that the class calls the built-in—in this case by importing the builtins
module which provides unambiguous access to all the built-in functions, and
calling builtins.ord() rather than plain ord().
Here’s another tiny yet complete class. This one allows us to create “constants”.
It isn’t difficult to change the values behind the class’s back, but it can at least
prevent simple mistakes.
class Const:
def __setattr__(self, name, value):
if name in self.__dict__:
raise ValueError("cannot change a const attribute")
self.__dict__[name] = value
def __delattr__(self, name):
if name in self.__dict__:
raise ValueError("cannot delete a const attribute")
raise AttributeError("'{0}' object has no attribute '{1}'"
.format(self.__class__.__name__, name))
With this class we can create a constant object, say, const = Const(), and set any
attributes we like on it, for example, const.limit = 591. But once an attribute’s
value has been set, although it can be read as often as we like, any attempt to
change or delete it will result in a ValueError exception being raised. We have
not reimplemented __getattr__() because the base class object.__getattr__()
method does what we want—returns the given attribute’s value or raises an
AttributeError exception if there is no such attribute. In the __delattr__()
method we mimic the __getattr__() method’s error message for nonexistent
attributes, and to do this we must get the name of the class we are in as well as
Further Object-Oriented Programming
Table 8.2 Attribute Access Special Methods
Special Method
__delattr__(self, name)
del x.n Deletes object x’s n attribute
Returns a list of x’s attribute
v = x.n Returns the value of object x’s n
attribute if it isn’t found directly
__getattr__(self, name)
__getattribute__(self, name)
v = x.n Returns the value of object x’s n
attribute; see text
__setattr__(self, name,
x.n = v Sets object x’s n attribute’s value
to v
the name of the nonexistent attribute. The class works because we are using
the object’s __dict__ which is what the base class __getattr__(), __setattr__(),
and __delattr__() methods use, although here we have used only the base
class’s __getattr__() method. All the special methods used for attribute access
are listed in Table 8.2.
There is another way of getting constants: We can use named tuples. Here are
a couple of examples:
Const = collections.namedtuple("_", "min max")(191, 591)
Const.min, Const.max
# returns: (191, 591)
Offset = collections.namedtuple("_", "id name description")(*range(3)),, Offset.description # returns: (0, 1, 2)
In both cases we have just used a throwaway name for the named tuple because we want just one named tuple instance each time, not a tuple subclass
for creating instances of a named tuple. Although Python does not support an
enum data type, we can use named tuples as we have done here to get a similar
261 ➤
For our last look at attribute access special methods we will return to an
example we first saw in Chapter 6. In that chapter we created an Image class
whose width, height, and background color are fixed when an Image is created
(although they are changed if an image is loaded). We provided access to them
using read-only properties. For example, we had:
def width(self):
return self.__width
This is easy to code but could become tedious if there are a lot of read-only
properties. Here is a different solution that handles all the Image class’s
read-only properties in a single method:
Chapter 8. Advanced Programming Techniques
def __getattr__(self, name):
if name == "colors":
return set(self.__colors)
classname = self.__class__.__name__
if name in frozenset({"background", "width", "height"}):
return self.__dict__["_{classname}__{name}".format(
raise AttributeError("'{classname}' object has no "
"attribute '{name}'".format(**locals()))
If we attempt to access an object’s attribute and the attribute is not found,
Python will call the __getattr__() method (providing it is implemented, and
that we have not reimplemented __getattribute__()), with the name of the
attribute as a parameter. Implementations of __getattr__() must raise an
AttributeError exception if they do not handle the given attribute.
For example, if we have the statement image.colors, Python will look for a colors attribute and having failed to find it, will then call Image.__getattr__(image,
"colors"). In this case the __getattr__() method handles a "colors" attribute
name and returns a copy of the set of colors that the image is using.
The other attributes are immutable, so they are safe to return directly to the
caller. We could have written separate elif statements for each one like this:
elif name == "background":
return self.__background
But instead we have chosen a more compact approach. Since we know that
under the hood all of an object’s nonspecial attributes are held in self.__dict__,
we have chosen to access them directly. For private attributes (those whose
name begins with two leading underscores), the name is mangled to have the
form _className__attributeName, so we must account for this when retrieving
the attribute’s value from the object’s private dictionary.
For the name mangling needed to look up private attributes and to provide the
standard AttributeError error text, we need to know the name of the class we
are in. (It may not be Image because the object might be an instance of an Image
subclass.) Every object has a __class__ special attribute, so self.__class__ is
always available inside methods and can safely be accessed by __getattr__()
without risking unwanted recursion.
Note that there is a subtle difference in that using __getattr__() and
self.__class__ provides access to the attribute in the instance’s class (which
may be a subclass), but accessing the attribute directly uses the class the attribute is defined in.
One special method that we have not covered is __getattribute__(). Whereas the __getattr__() method is called last when looking for (nonspecial)
Chapter 8. Advanced Programming Techniques
def make_strip_function(characters):
def strip_function(string):
return string.strip(characters)
return strip_function
strip_punctuation = make_strip_function(",;:.!?")
strip_punctuation("Land ahoy!")
# returns: 'Land ahoy'
The make_strip_function() function takes the characters to be stripped as its
sole argument and returns a function, strip_function(), that takes a string
argument and which strips the characters that were given at the time the
closure was created. So just as we can create as many instances of the Strip
class as we want, each with its own characters to strip, we can create as many
strip functions with their own characters as we like.
The classic use case for functors is to provide key functions for sort routines.
Here is a generic SortKey functor class (from file
class SortKey:
def __init__(self, *attribute_names):
self.attribute_names = attribute_names
def __call__(self, instance):
values = []
for attribute_name in self.attribute_names:
values.append(getattr(instance, attribute_name))
return values
When a SortKey object is created it keeps a tuple of the attribute names it
was initialized with. When the object is called it creates a list of the attribute
values for the instance it is passed—in the order they were specified when the
SortKey was initialized. For example, imagine we have a Person class:
class Person:
def __init__(self, forename, surname, email):
self.forename = forename
self.surname = surname = email
Suppose we have a list of Person objects in the people list. We can sort the
list by surnames like this: people.sort(key=SortKey("surname")). If there
are a lot of people there are bound to be some surname clashes, so we can
sort by surname, and then by forename within surname, like this: people.sort(key=SortKey("surname", "forename")). And if we had people with the
same surname and forename we could add the email attribute too. And of
Further Object-Oriented Programming
course, we could sort by forename and then surname by changing the order of
the attribute names we give to the SortKey functor.
Another way of achieving the same thing, but without needing to create a functor at all, is to use the operator module’s operator.attrgetter() function. For
example, to sort by surname we could write: people.sort(key=operator.attrgetter("surname")). And similarly, to sort by surname and forename:
people.sort(key=operator.attrgetter("surname", "forename")). The operator.
attrgetter() function returns a function (a closure) that, when called on an object, returns those attributes of the object that were specified when the closure
was created.
Functors are probably used rather less frequently in Python than in other
languages that support them because Python has other means of doing the
same things—for example, using closures or item and attribute getters.
Context Managers
Context managers allow us to simplify code by ensuring that certain operations are performed before and after a particular block of code is executed. The
behavior is achieved because context managers define two special methods,
__enter__() and __exit__(), that Python treats specially in the scope of a with
statement. When a context manager is created in a with statement its __enter__() method is automatically called, and when the context manager goes out
of scope after its with statement its __exit__() method is automatically called.
We can create our own custom context managers or use predefined ones—as
we will see later in this subsection, the file objects returned by the built-in
open() function are context managers. The syntax for using context managers
is this:
with expression as variable:
The expression must be or must produce a context manager object; if the
optional as variable part is specified, the variable is set to refer to the object
returned by the context manager’s __enter__() method (and this is often the
context manager itself). Because a context manager is guaranteed to execute
its “exit” code (even in the face of exceptions), context managers can be used to
eliminate the need for finally blocks in many situations.
Some of Python’s types are context managers—for example, all the file objects
that open() can return—so we can eliminate finally blocks when doing file
handling as these equivalent code snippets illustrate (assuming that process()
is a function defined elsewhere):
Chapter 8. Advanced Programming Techniques
fh = None
fh = open(filename)
for line in fh:
except EnvironmentError as err:
if fh is not None:
with open(filename) as fh:
for line in fh:
except EnvironmentError as err:
A file object is a context manager whose exit code always closes the file if it
was opened. The exit code is executed whether or not an exception occurs, but
in the latter case, the exception is propagated. This ensures that the file gets
closed and we still get the chance to handle any errors, in this case by printing
a message for the user.
In fact, context managers don’t have to propagate exceptions, but not doing so
effectively hides any exceptions, and this would almost certainly be a coding
error. All the built-in and standard library context managers propagate exceptions.
Sometimes we need to use more than one context manager at the same time.
For example:
with open(source) as fin:
with open(target, "w") as fout:
for line in fin:
except EnvironmentError as err:
Here we read lines from the source file and write processed versions of them to
the target file.
Using nested with statements can quickly lead to a lot of indentation. Fortunately, the standard library’s contextlib module provides some additional support for context managers, including the contextlib.nested() function which
allows two or more context managers to be handled in the same with statement
rather than having to nest with statements. Here is a replacement for the code
just shown, but omitting most of the lines that are identical to before:
with contextlib.nested(open(source), open(target, "w")) as (
fin, fout):
for line in fin:
Further Object-Oriented Programming
It is only necessary to use contextlib.nested() for Python 3.0; from Python 3.1
this function is deprecated because Python 3.1 can handle multiple context
managers in a single with statement. Here is the same example—again
omitting irrelevant lines—but this time for Python 3.1:
with open(source) as fin, open(target, "w") as fout:
for line in fin:
Using this syntax keeps context managers and the variables they are associated with together, making the with statement much more readable than if we
were to nest them or to use contextlib.nested().
It isn’t only file objects that are context managers. For example, several
threading-related classes used for locking are context managers. Context
managers can also be used with decimal.Decimal numbers; this is useful if we
want to perform some calculations with certain settings (such as a particular
precision) in effect.
If we want to create a custom context manager we must create a class that
provides two methods: __enter__() and __exit__(). Whenever a with statement
is used on an instance of such a class, the __enter__() method is called and
the return value is used for the as variable (or thrown away if there isn’t one).
When control leaves the scope of the with statement the __exit__() method is
called (with details of an exception if one has occurred passed as arguments).
Suppose we want to perform several operations on a list in an atomic
manner—that is, we either want all the operations to be done or none of them
so that the resultant list is always in a known state. For example, if we have
a list of integers and want to append an integer, delete an integer, and change
a couple of integers, all as a single operation, we could write code like this:
with AtomicList(items) as atomic:
del atomic[3]
atomic[8] = 81738
atomic[index] = 38172
except (AttributeError, IndexError, ValueError) as err:
print("no changes applied:", err)
If no exception occurs, all the operations are applied to the original list (items),
but if an exception occurs, no changes are made at all. Here is the code for the
AtomicList context manager:
class AtomicList:
def __init__(self, alist, shallow_copy=True):
➤ 439
Chapter 8. Advanced Programming Techniques
self.original = alist
self.shallow_copy = shallow_copy
def __enter__(self):
self.modified = (self.original[:] if self.shallow_copy
else copy.deepcopy(self.original))
return self.modified
def __exit__(self, exc_type, exc_val, exc_tb):
if exc_type is None:
self.original[:] = self.modified
146 ➤
When the AtomicList object is created we keep a reference to the original list
and note whether shallow copying is to be used. (Shallow copying is fine for
lists of numbers or strings; but for lists that contain lists or other collections,
shallow copying is not sufficient.)
Then, when the AtomicList context manager object is used in the with statement its __enter__() method is called. At this point we copy the original list
and return the copy so that all the changes can be made on the copy.
Once we reach the end of the with statement’s scope the __exit__() method is
called. If no exception occurred the exc_type (“exception type”) will be None and
we know that we can safely replace the original list’s items with the items from
the modified list. (We cannot do self.original = self.modified because that
would just replace one object reference with another and would not affect the
original list at all.) But if an exception occurred, we do nothing to the original
list and the modified list is discarded.
The return value of __exit__() is used to indicate whether any exception that
occurred should be propagated. A True value means that we have handled any
exception and so no propagation should occur. Normally we always return
False or something that evaluates to False in a Boolean context to allow any
exception that occurred to propagate. By not giving an explicit return value,
our __exit__() returns None which evaluates to False and correctly causes any
exception to propagate.
Custom context managers are used in Chapter 11 to ensure that socket
connections and gzipped files are closed, and some of the threading modules
context managers are used in Chapter 10 to ensure that mutual exclusion locks
are unlocked. You’ll also get the chance to create a more generic atomic contex
manager in this chapter’s exercises.
Descriptors are classes which provide access control for the attributes of other
classes. Any class that implements one or more of the descriptor special
Further Object-Oriented Programming
methods, __get__(), __set__(), and __delete__(), is called (and can be used as)
a descriptor.
The built-in property() and classmethod() functions are implemented using
descriptors. The key to understanding descriptors is that although we create
an instance of a descriptor in a class as a class attribute, Python accesses the
descriptor through the class’s instances.
To make things clear, let’s imagine that we have a class whose instances hold
some strings. We want to access the strings in the normal way, for example,
as a property, but we also want to get an XML-escaped version of the strings
whenever we want. One simple solution would be that whenever a string is
set we immediately create an XML-escaped copy. But if we had thousands
of strings and only ever read the XML version of a few of them, we would
be wasting a lot of processing and memory for nothing. So we will create a
descriptor that will provide XML-escaped strings on demand without storing
them. We will start with the beginning of the client (owner) class, that is, the
class that uses the descriptor:
class Product:
__slots__ = ("__name", "__description", "__price")
name_as_xml = XmlShadow("name")
description_as_xml = XmlShadow("description")
def __init__(self, name, description, price):
self.__name = name
self.description = description
self.price = price
The only code we have not shown are the properties; the name is a read-only
property and the description and price are readable/writable properties, all set
up in the usual way. (All the code is in the file.) We have used the
__slots__ variable to ensure that the class has no __dict__ and can store only
the three specified private attributes; this is not related to or necessary for our
use of descriptors. The name_as_xml and description_as_xml class attributes are
set to be instances of the XmlShadow descriptor. Although no Product object has a
name_as_xml attribute or a description_as_xml attribute, thanks to the descriptor
we can write code like this (here quoting from the module’s doctests):
>>> product = Product("Chisel <3cm>", "Chisel & cap", 45.25)
>>>, product.name_as_xml, product.description_as_xml
('Chisel <3cm>', 'Chisel &lt;3cm&gt;', 'Chisel &amp; cap')
This works because when we try to access, for example, the name_as_xml
attribute, Python finds that the Product class has a descriptor with that name,
Chapter 8. Advanced Programming Techniques
and so uses the descriptor to get the attribute’s value. Here’s the complete code
for the XmlShadow descriptor class:
class XmlShadow:
def __init__(self, attribute_name):
self.attribute_name = attribute_name
def __get__(self, instance, owner=None):
return xml.sax.saxutils.escape(
getattr(instance, self.attribute_name))
When the name_as_xml and description_as_xml objects are created we pass the
name of the Product class’s corresponding attribute to the XmlShadow initializer so that the descriptor knows which attribute to work on. Then, when the
name_as_xml or description_as_xml attribute is looked up, Python calls the descriptor’s __get__() method. The self argument is the instance of the descriptor, the instance argument is the Product instance (i.e., the product’s self), and
the owner argument is the owning class (Product in this case). We use the getattr() function to retrieve the relevant attribute from the product (in this case
the relevant property), and return an XML-escaped version of it.
If the use case was that only a small proportion of the products were accessed
for their XML strings, but the strings were often long and the same ones were
frequently accessed, we could use a cache. For example:
class CachedXmlShadow:
def __init__(self, attribute_name):
self.attribute_name = attribute_name
self.cache = {}
def __get__(self, instance, owner=None):
xml_text = self.cache.get(id(instance))
if xml_text is not None:
return xml_text
return self.cache.setdefault(id(instance),
getattr(instance, self.attribute_name)))
We store the unique identity of the instance as the key rather than the instance
itself because dictionary keys must be hashable (which IDs are), but we don’t
want to impose that as a requirement on classes that use the CachedXmlShadow descriptor. The key is necessary because descriptors are created per class
rather than per instance. (The dict.setdefault() method conveniently returns
the value for the given key, or if no item with that key is present, creates a new
item with the given key and value and returns the value.)
Further Object-Oriented Programming
Having seen descriptors used to generate data without necessarily storing it,
we will now look at a descriptor that can be used to store all of an object’s attribute data, with the object not needing to store anything itself. In the example, we will just use a dictionary, but in a more realistic context, the data might
be stored in a file or a database. Here’s the start of a modified version of the
Point class that makes use of the descriptor (from the file):
class Point:
__slots__ = ()
x = ExternalStorage("x")
y = ExternalStorage("y")
def __init__(self, x=0, y=0):
self.x = x
self.y = y
By setting __slots__ to an empty tuple we ensure that the class cannot store
any data attributes at all. When self.x is assigned to, Python finds that there
is a descriptor with the name “x”, and so uses the descriptor’s __set__() method.
The rest of the class isn’t shown, but is the same as the original Point class
shown in Chapter 6. Here is the complete ExternalStorage descriptor class:
class ExternalStorage:
__slots__ = ("attribute_name",)
__storage = {}
def __init__(self, attribute_name):
self.attribute_name = attribute_name
def __set__(self, instance, value):
self.__storage[id(instance), self.attribute_name] = value
def __get__(self, instance, owner=None):
if instance is None:
return self
return self.__storage[id(instance), self.attribute_name]
Each ExternalStorage object has a single data attribute, attribute_name, which
holds the name of the owner class’s data attribute. Whenever an attribute
is set we store its value in the private class dictionary, __storage. Similarly,
whenever an attribute is retrieved we get it from the __storage dictionary.
As with all descriptor methods, self is the instance of the descriptor object and
instance is the self of the object that contains the descriptor, so here self is an
ExternalStorage object and instance is a Point object.
Chapter 8. Advanced Programming Techniques
Although __storage is a class attribute, we can access it as self.__storage (just
as we can call methods using self.method()), because Python will look for it as
an instance attribute, and not finding it will then look for it as a class attribute.
The one (theoretical) disadvantage of this approach is that if we have a class
attribute and an instance attribute with the same name, one would hide the
other. (If this were really a problem we could always refer to the class attribute
using the class, that is, ExternalStorage.__storage. Although hard-coding the
class does not play well with subclassing in general, it doesn’t really matter
for private attributes since Python name-mangles the class name into them
The implementation of the __get__() special method is slightly more sophisticated than before because we provide a means by which the ExternalStorage
instance itself can be accessed. For example, if we have p = Point(3, 4), we can
access the x-coordinate with p.x, and we can access the ExternalStorage object
that holds all the xs with Point.x.
To complete our coverage of descriptors we will create the Property descriptor
that mimics the behavior of the built-in property() function, at least for setters
and getters. The code is in Here is the complete NameAndExtension
class that makes use of it:
class NameAndExtension:
def __init__(self, name, extension):
self.__name = name
self.extension = extension
def name(self):
return self.__name
# Uses the custom Property descriptor
# Uses the custom Property descriptor
def extension(self):
return self.__extension
# Uses the custom Property descriptor
def extension(self, extension):
self.__extension = extension
The usage is just the same as for the built-in @property decorator and for the
@propertyName.setter decorator. Here is the start of the Property descriptor’s
class Property:
def __init__(self, getter, setter=None):
self.__getter = getter
Further Object-Oriented Programming
self.__setter = setter
self.__name__ = getter.__name__
The class’s initializer takes one or two functions as arguments. If it is used as
a decorator, it will get just the decorated function and this becomes the getter,
while the setter is set to None. We use the getter’s name as the property’s name.
So for each property, we have a getter, possibly a setter, and a name.
def __get__(self, instance, owner=None):
if instance is None:
return self
return self.__getter(instance)
When a property is accessed we return the result of calling the getter function where we have passed the instance as its first parameter. At first sight,
self.__getter() looks like a method call, but it is not. In fact, self.__getter
is an attribute, one that happens to hold an object reference to a method
that was passed in. So what happens is that first we retrieve the attribute
(self.__getter), and then we call it as a function (). And because it is called as
a function rather than as a method we must pass in the relevant self object
explicitly ourselves. And in the case of a descriptor the self object (from the
class that is using the descriptor) is called instance (since self is the descriptor
object). The same applies to the __set__() method.
def __set__(self, instance, value):
if self.__setter is None:
raise AttributeError("'{0}' is read-only".format(
return self.__setter(instance, value)
If no setter has been specified, we raise an AttributeError; otherwise, we call
the setter with the instance and the new value.
def setter(self, setter):
self.__setter = setter
return self.__setter
This method is called when the interpreter reaches, for example, @extension.setter, with the function it decorates as its setter argument. It stores
the setter method it has been given (which can now be used in the __set__()
method), and returns the setter, since decorators should return the function or
method they decorate.
We have now looked at three quite different uses of descriptors. Descriptors
are a very powerful and flexible feature that can be used to do lots of underthe-hood work while appearing to be simple attributes in their client (owner) class.
Chapter 8. Advanced Programming Techniques
Class Decorators
Just as we can create decorators for functions and methods, we can also create
decorators for entire classes. Class decorators take a class object (the result of
the class statement), and should return a class—normally a modified version
of the class they decorate. In this subsection we will study two class decorators
to see how they can be implemented.
269 ➤
In Chapter 6 we created the SortedList custom collection class that aggregated
a plain list as the private attribute self.__list. Eight of the SortedList methods simply passed on their work to the private attribute. For example, here are
how the SortedList.clear() and SortedList.pop() methods were implemented:
def clear(self):
self.__list = []
def pop(self, index=-1):
return self.__list.pop(index)
There is nothing we can do about the clear() method since there is no corresponding method for the list type, but for pop(), and the other six methods that
SortedList delegates, we can simply call the list class’s corresponding method.
This can be done by using the @delegate class decorator from the book’s Util
module. Here is the start of a new version of the SortedList class:
@Util.delegate("__list", ("pop", "__delitem__", "__getitem__",
"__iter__", "__reversed__", "__len__", "__str__"))
class SortedList:
The first argument is the name of the attribute to delegate to, and the second
argument is a sequence of one or more methods that we want the delegate()
decorator to implement for us so that we don’t have to do the work ourselves.
The SortedList class in the file uses this approach and
therefore does not have any code for the methods listed, even though it fully
supports them. Here is the class decorator that implements the methods:
def delegate(attribute_name, method_names):
def decorator(cls):
nonlocal attribute_name
if attribute_name.startswith("__"):
attribute_name = "_" + cls.__name__ + attribute_name
for name in method_names:
setattr(cls, name, eval("lambda self, *a, **kw: "
"self.{0}.{1}(*a, **kw)".format(
attribute_name, name)))
return cls
return decorator
Further Object-Oriented Programming
We could not use a plain decorator because we want to pass arguments to the
decorator, so we have instead created a function that takes our arguments and
that returns a class decorator. The decorator itself takes a single argument,
a class (just as a function decorator takes a single function or method as
its argument).
We must use nonlocal so that the nested function uses the attribute_name from
the outer scope rather than attempting to use one from its own scope. And
we must be able to correct the attribute name if necessary to take account of
the name mangling of private attributes. The decorator’s behavior is quite
simple: It iterates over all the method names that the delegate() function has
been given, and for each one creates a new method which it sets as an attribute
on the class with the given method name.
We have used eval() to create each of the delegated methods since it can be
used to execute a single statement, and a lambda statement produces a method
or function. For example, the code executed to produce the pop() method is:
lambda self, *a, **kw: self._SortedList__list.pop(*a, **kw)
We use the * and ** argument forms to allow for any arguments even though
the methods being delegated to have specific argument lists. For example,
list.pop() accepts a single index position (or nothing, in which case it defaults
to the last item). This is okay because if the wrong number or kinds of arguments are passed, the list method that is called to do the work will raise an
appropriate exception.
248 ➤
The second class decorator we will review was also used in Chapter 6. When
we implemented the FuzzyBool class we mentioned that we had supplied only
the __lt__() and __eq__() special methods (for < and ==), and had generated all
the other comparison methods automatically. What we didn’t show was the
complete start of the class definition:
class FuzzyBool:
The other four comparison operators were provided by the complete_comparisons() class decorator. Given a class that defines only < (or < and ==), the decorator produces the missing comparison operators by using the following logical
¬ (x < y ∨ y < x)
¬ (x = y)
¬ (y < x)
¬ (x < y)
If the class to be decorated has < and ==, the decorator will use them both,
falling back to doing everything in terms of < if that is the only operator
Chapter 8. Advanced Programming Techniques
supplied. (In fact, Python automatically produces > if < is supplied, != if == is
supplied, and >= if <= is supplied, so it is sufficient to just implement the three
operators <, <=, and == and to leave Python to infer the others. However, using
the class decorator reduces the minimum that we must implement to just <.
This is convenient, and also ensures that all the comparison operators use the
same consistent logic.)
def complete_comparisons(cls):
assert cls.__lt__ is not object.__lt__, (
"{0} must define < and ideally ==".format(cls.__name__))
if cls.__eq__ is object.__eq__:
cls.__eq__ = lambda self, other: (not
(cls.__lt__(self, other) or cls.__lt__(other, self)))
cls.__ne__ = lambda self, other: not cls.__eq__(self, other)
cls.__gt__ = lambda self, other: cls.__lt__(other, self)
cls.__le__ = lambda self, other: not cls.__lt__(other, self)
cls.__ge__ = lambda self, other: not cls.__lt__(self, other)
return cls
One problem that the decorator faces is that class object from which every
other class is ultimately derived defines all six comparison operators, all of
which raise a TypeError exception if used. So we need to know whether < and
== have been reimplemented (and are therefore usable). This can easily be done
by comparing the relevant special methods in the class being decorated with
those in object.
If the decorated class does not have a custom < the assertion fails because that
is the decorator’s minimum requirement. And if there is a custom == we use
it; otherwise, we create one. Then all the other methods are created and the
decorated class, now with all six comparison methods, is returned.
Using class decorators is probably the simplest and most direct way of
changing classes. Another approach is to use metaclasses, a topic we will cover
later in this chapter.
Abstract Base Classes
An abstract base class (ABC) is a class that cannot be used to create objects.
Instead, the purpose of such classes is to define interfaces, that is, to in effect
list the methods and properties that classes that inherit the abstract base class
must provide. This is useful because we can use an abstract base class as a
kind of promise—a promise that any derived class will provide the methods
and properties that the abstract base class specifies.★
Python’s abstract base classes are described in PEP 3119 (,
which also includes a very useful rationale and is well worth reading.
➤ 390
Further Object-Oriented Programming
Table 8.3 The Numbers Module’s Abstract Base Classes
==, !=, +, -, *, /, abs(), bool(),
complex(), conjugate(); also real
and imag properties
<, <=, ==, !=, >=, >, +, -, *, /,
//, %, abs(), bool(), complex(),
conjugate(), divmod(), float(),
math.ceil(), math.floor(), round(), int
trunc(); also real and imag
Rational <, <=, ==, !=, >=, >, +, -, *, /, //,
%, <<, >>, ~, &, ^, |, abs(), bool(),
complex(), conjugate(), divmod(),
float(), math.ceil(), math.floor(),
pow(), round(), trunc(); also real,
imag, numerator, and denominator
<, <=, ==, !=, >=, >, +, -, *, /,
//, %, abs(), bool(), complex(),
conjugate(), divmod(), float(),
math.ceil(), math.floor(), round(),
trunc(); also real, imag, numerator,
and denominator properties
Abstract base classes are classes that have at least one abstract method or
property. Abstract methods can be defined with no implementation (i.e., their
suite is pass, or if we want to force reimplementation in a subclass, raise
NotImplementedError()), or with an actual (concrete) implementation that can
be invoked from subclasses, for example, when there is a common case. They
can also have other concrete (i.e., nonabstract) methods and properties.
Classes that derive from an ABC can be used to create instances only if they
reimplement all the abstract methods and abstract properties they have inherited. For those abstract methods that have concrete implementations (even if
it is only pass), the derived class could simply use super() to use the ABC’s
Chapter 8. Advanced Programming Techniques
sion. Any concrete methods or properties are available through inheritance as
usual. All ABCs must have a metaclass of abc.ABCMeta (from the abc module),
or from one of its subclasses. We cover metaclasses a bit further on.
➤ 390
Python provides two groups of abstract base classes, one in the collections
module and the other in the numbers module. They allow us to ask questions
about an object; for example, given a variable x, we can see whether it is a sequence using isinstance(x, collections.MutableSequence) or whether it is a
whole number using isinstance(x, numbers.Integral). This is particularly useful in view of Python’s dynamic typing where we don’t necessarily know (or
care) what an object’s type is, but want to know whether it supports the operations we want to apply to it. The numeric and collection ABCs are listed in
Tables 8.3 and 8.4. The other major ABC is io.IOBase from which all the file and
stream-handling classes derive.
To fully integrate our own custom numeric and collection classes we ought to
make them fit in with the standard ABCs. For example, the SortedList class is
a sequence, but as it stands, isinstance(L, collections.Sequence) returns False
if L is a SortedList. One easy way to fix this is to inherit the relevant ABC:
class SortedList(collections.Sequence):
By making collections.Sequence the base class, the isinstance() test will
now return True. Furthermore, we will be required to implement __init__()
(or __new__()), __getitem__(), and __len__() (which we do). The collections.Sequence ABC also provides concrete (i.e., nonabstract) implementations
for __contains__(), __iter__(), __reversed__(), count(), and index(). In the case
of SortedList, we reimplement them all, but we could have used the ABC versions if we wanted to, simply by not reimplementing them. We cannot make
SortedList a subclass of collections.MutableSequence even though the list
is mutable because SortedList does not have all the methods that a collections.MutableSequence must provide, such as __setitem__() and append(). (The
code for this SortedList is in We will see an alternative approach to making a SortedList into a collections.Sequence in the Metaclasses
Now that we have seen how to make a custom class fit in with the standard
ABCs, we will turn to another use of ABCs: to provide an interface promise
for our own custom classes. We will look at three rather different examples to
cover different aspects of creating and using ABCs.
We will start with a very simple example that shows how to handle readable/writable properties. The class is used to represent domestic appliances.
Every appliance that is created must have a read-only model string and a readable/writable price. We also want to ensure that the ABC’s __init__() is reimplemented. Here’s the ABC (from; we have not shown the import
➤ 390
Further Object-Oriented Programming
Table 8.4 The Collections Module’s Main Abstract Base Classes
All functions,
methods, and
bytearray, bytes,
dict, frozenset,
list, set, str, tuple
bytes, frozenset,
str, tuple
bytearray, bytes,
dict, frozenset,
list, set, str, tuple
iter(), next()
bytearray, bytes,
dict, frozenset,
list, set, str, tuple
==, !=, [], len(), iter(),
in, get(), items(), keys(),
==, !=, [], del, len(), iter(),
in, clear(), get(), items(),
keys(), pop(), popitem(),
setdefault(), update(),
[], len(), iter(), reversed(),
in, count(), index()
bytearray, bytes,
list, str, tuple
[], +=, del, len(), iter(),
reversed(), in, append(),
count(), extend(), index(),
insert(), pop(), remove(),
bytearray, list
<, <=, ==, !=, =>, >, &, |, ^, len(), frozenset, set
iter(), in, isdisjoint()
<, <=, ==, !=, =>, >, &, |, ^,
&=, |=, ^=, -=, len(), iter(),
in, add(), clear(), discard(),
isdisjoint(), pop(), remove()
Chapter 8. Advanced Programming Techniques
abc statement which is needed for the abstractmethod() and abstractproperty()
functions, both of which can be used as decorators:
class Appliance(metaclass=abc.ABCMeta):
def __init__(self, model, price):
self.__model = model
self.price = price
def get_price(self):
return self.__price
def set_price(self, price):
self.__price = price
price = abc.abstractproperty(get_price, set_price)
def model(self):
return self.__model
We have set the class’s metaclass to be abc.ABCMeta since this is a requirement
for ABCs; any abc.ABCMeta subclass can be used instead, of course. We have
made __init__() an abstract method to ensure that it is reimplemented, and
we have also provided an implementation which we expect (but can’t force)
inheritors to call. To make an abstract readable/writable property we cannot
use decorator syntax; also we have not used private names for the getter and
setter since doing so would be inconvenient for subclasses.
The price property is abstract (so we cannot use the @property decorator), and is
readable/writable. Here we follow a common pattern for when we have private
readable/writable data (e.g., __price) as a property: We initialize the property
in the __init__() method rather than setting the private data directly—this
ensures that the setter is called (and may potentially do validation or other
work, although it doesn’t in this particular example).
The model property is not abstract, so subclasses don’t need to reimplement it,
and we can make it a property using the @property decorator. Here we follow
a common pattern for when we have private read–only data (e.g., __model) as
a property: We set the private __model data once in the __init__() method, and
provide read access via the read–only model property.
Note that no Appliance objects can be created, because the class contains
abstract attributes. Here is an example subclass:
Further Object-Oriented Programming
class Cooker(Appliance):
def __init__(self, model, price, fuel):
super().__init__(model, price)
self.fuel = fuel
price = property(lambda self: super().price,
lambda self, price: super().set_price(price))
The Cooker class must reimplement the __init__() method and the price
property. For the property we have just passed on all the work to the base class.
The model read-only property is inherited. We could create many more classes
based on Appliance, such as Fridge, Toaster, and so on.
The next ABC we will look at is even shorter; it is an ABC for text-filtering
functors (in file
class TextFilter(metaclass=abc.ABCMeta):
def is_transformer(self):
raise NotImplementedError()
def __call__(self):
raise NotImplementedError()
The TextFilter ABC provides no functionality at all; it exists purely to define
an interface, in this case an is_transformer read-only property and a __call__()
method, that all its subclasses must provide. Since the abstract property and
method have no implementations we don’t want subclasses to call them, so
instead of using an innocuous pass statement we raise an exception if they are
used (e.g., via a super() call).
Here is one simple subclass:
class CharCounter(TextFilter):
def is_transformer(self):
return False
def __call__(self, text, chars):
count = 0
for c in text:
if c in chars:
count += 1
return count
Chapter 8. Advanced Programming Techniques
This text filter is not a transformer because rather than transforming the text
it is given, it simply returns a count of the specified characters that occur in
the text. Here is an example of use:
vowel_counter = CharCounter()
vowel_counter("dog fish and cat fish", "aeiou")
# returns: 5
Two other text filters are provided, both of which are transformers: RunLengthEncode and RunLengthDecode. Here is how they are used:
rle_encoder = RunLengthEncode()
rle_text = rle_encoder(text)
rle_decoder = RunLengthDecode()
original_text = rle_decoder(rle_text)
The run length encoder converts a string into UTF-8 encoded bytes, and
replaces 0x00 bytes with the sequence 0x00, 0x01, 0x00, and any sequence of
three to 255 repeated bytes with the sequence 0x00, count, byte. If the string has
lots of runs of four or more identical consecutive characters this can produce a
shorter byte string than the raw UTF-8 encoded bytes. The run length decoder
takes a run length encoded byte string and returns the original string. Here is
the start of the RunLengthDecode class:
class RunLengthDecode(TextFilter):
def is_transformer(self):
return True
def __call__(self, rle_bytes):
We have omitted the body of the __call__() method, although it is in the
source that accompanies this book. The RunLengthEncode class has exactly the
same structure.
The last ABC we will look at provides an Application Programming Interface
(API) and a default implementation for an undo mechanism. Here is the
complete ABC (from file
class Undo(metaclass=abc.ABCMeta):
def __init__(self):
self.__undos = []
Further Object-Oriented Programming
def can_undo(self):
return bool(self.__undos)
def undo(self):
assert self.__undos, "nothing left to undo"
def add_undo(self, undo):
The __init__() and undo() methods must be reimplemented since they are
both abstract; and so must the read-only can_undo property. Subclasses don’t
have to reimplement the add_undo() method, although they are free to do so.
The undo() method is slightly subtle. The self.__undos list is expected to hold
object references to methods. Each method must cause the corresponding
action to be undone if it is called—this will be clearer when we look at an Undo
subclass in a moment. So to perform an undo we pop the last undo method off
the self.__undos list, and then call the method as a function, passing self as an
argument. (We must pass self because the method is being called as a function
and not as a method.)
Here is the beginning of the Stack class; it inherits Undo, so any actions performed on it can be undone by calling Stack.undo() with no arguments:
class Stack(Undo):
def __init__(self):
self.__stack = []
def can_undo(self):
return super().can_undo
def undo(self):
def push(self, item):
self.add_undo(lambda self: self.__stack.pop())
def pop(self):
item = self.__stack.pop()
self.add_undo(lambda self: self.__stack.append(item))
return item
Chapter 8. Advanced Programming Techniques
We have omitted and Stack.__str__() since neither adds anything
new and neither interacts with the Undo base class. For the can_undo property
and the undo() method, we simply pass on the work to the base class. If these
two were not abstract we would not need to reimplement them at all and the
same effect would be achieved; but in this case we wanted to force subclasses
to reimplement them to encourage undo to be taken account of in the subclass.
For push() and pop() we perform the operation and also add a function to the
undo list which will undo the operation that has just been performed.
Abstract base classes are most useful in large-scale programs, libraries, and
application frameworks, where they can help ensure that irrespective of
implementation details or author, classes can work cooperatively together
because they provide the APIs that their ABCs specify.
Multiple Inheritance
Multiple inheritance is where one class inherits from two or more other classes.
Although Python (and, for example, C++) fully supports multiple inheritance,
some languages—most notably, Java—don’t allow it. One problem is that
multiple inheritance can lead to the same class being inherited more than once
(e.g., if two of the base classes inherit from the same class), and this means that
the version of a method that is called, if it is not in the subclass but is in two
or more of the base classes (or their base classes, etc.), depends on the method
resolution order, which potentially makes classes that use multiple inheritance
somewhat fragile.
Multiple inheritance can generally be avoided by using single inheritance (one
base class), and setting a metaclass if we want to support an additional API,
since as we will see in the next subsection, a metaclass can be used to give the
promise of an API without actually inheriting any methods or data attributes.
An alternative is to use multiple inheritance with one concrete class and one
or more abstract base classes for additional APIs. And another alternative is
to use single inheritance and aggregate instances of other classes.
Nonetheless, in some cases, multiple inheritance can provide a very convenient
solution. For example, suppose we want to create a new version of the Stack
class from the previous subsection, but want the class to support loading and
saving using a pickle. We might well want to add the loading and saving
functionality to several classes, so we will implement it in a class of its own:
class LoadSave:
def __init__(self, filename, *attribute_names):
self.filename = filename
self.__attribute_names = []
for name in attribute_names:
if name.startswith("__"):
Further Object-Oriented Programming
name = "_" + self.__class__.__name__ + name
def save(self):
with open(self.filename, "wb") as fh:
data = []
for name in self.__attribute_names:
data.append(getattr(self, name))
pickle.dump(data, fh, pickle.HIGHEST_PROTOCOL)
def load(self):
with open(self.filename, "rb") as fh:
data = pickle.load(fh)
for name, value in zip(self.__attribute_names, data):
setattr(self, name, value)
The class has two attributes: filename, which is public and can be changed at
any time, and __attribute_names, which is fixed and can be set only when the
instance is created. The save() method iterates over all the attribute names
and creates a list called data that holds the value of each attribute to be saved;
it then saves the data into a pickle. The with statement ensures that the file is
closed if it was successfully opened, and any file or pickle exceptions are passed
up to the caller. The load() method iterates over the attribute names and the
corresponding data items that have been loaded and sets each attribute to its
loaded value.
Here is the start of the FileStack class that multiply-inherits the Undo class
from the previous subsection and this subsection’s LoadSave class:
class FileStack(Undo, LoadSave):
def __init__(self, filename):
LoadSave.__init__(self, filename, "__stack")
self.__stack = []
def load(self):
The rest of the class is just the same as the Stack class, so we have not reproduced it here. Instead of using super() in the __init__() method we must specify the base classes that we initialize since super() cannot guess our intentions.
For the LoadSave initialization we pass the filename to use and also the names
of the attributes we want saved; in this case just one, the private __stack. (We
don’t want to save the __undos; and nor could we in this case since it is a list of
methods and is therefore unpicklable.)
Chapter 8. Advanced Programming Techniques
The FileStack class has all the Undo methods, and also the LoadSave class’s save()
and load() methods. We have not reimplemented save() since it works fine,
but for load() we must clear the undo stack after loading. This is necessary
because we might do a save, then do various changes, and then a load. The load
wipes out what went before, so any undos no longer make sense. The original
Undo class did not have a clear() method, so we had to add one:
def clear(self):
self.__undos = []
# In class Undo
In the Stack.load() method we have used super() to call LoadSave.load() because there is no Undo.load() method to cause ambiguity. If both base classes had had a load() method, the one that would get called would depend on
Python’s method resolution order. We prefer to use super() only when there
is no ambiguity, and to use the appropriate base name otherwise, so we never
rely on the method resolution order. For the self.clear() call, again there is no
ambiguity since only the Undo class has a clear() method, and we don’t need to
use super() since (unlike load()) FileStack does not have a clear() method.
What would happen if, later on, a clear() method was added to the FileStack
class? It would break the load() method. One solution would be to call super().clear() inside load() instead of plain self.clear(). This would result in
the first super-class’s clear() method that was found being used. To protect
against such problems we could make it a policy to use hard-coded base classes
when using multiple inheritance (in this example, calling Undo.clear(self)). Or
we could avoid multiple inheritance altogether and use aggregation, for example, inheriting the Undo class and creating a LoadSave class designed for aggregation.
What multiple inheritance has given us here is a mixture of two rather different classes, without the need to implement any of the undo or the loading
and saving ourselves, relying instead on the functionality provided by the base
classes. This can be very convenient and works especially well when the inherited classes have no overlapping APIs.
A metaclass is to a class what a class is to an instance; that is, a metaclass is
used to create classes, just as classes are used to create instances. And just as
we can ask whether an instance belongs to a class by using isinstance(), we
can ask whether a class object (such as dict, int, or SortedList) inherits another
class using issubclass().
The simplest use of metaclasses is to make custom classes fit into Python’s
standard ABC hierarchy. For example, to make SortedList a collections.
Further Object-Oriented Programming
Sequence, instead of inheriting the ABC (as we showed earlier), we can simply
register the SortedList as a collections.Sequence:
class SortedList:
After the class is defined normally, we register it with the collections.Sequence
ABC. Registering a class like this makes it a virtual subclass.★ A virtual subclass reports that it is a subclass of the class or classes it is registered with (e.g.,
using isinstance() or issubclass()), but does not inherit any data or methods
from any of the classes it is registered with.
Registering a class like this provides a promise that the class provides the API
of the classes it is registered with, but does not provide any guarantee that it
will honor its promise. One use of metaclasses is to provide both a promise and
a guarantee about a class’s API. Another use is to modify a class in some way
(like a class decorator does). And of course, metaclasses can be used for both
purposes at the same time.
Suppose we want to create a group of classes that all provide load() and save()
methods. We can do this by creating a class that when used as a metaclass,
checks that these methods are present:
class LoadableSaveable(type):
def __init__(cls, classname, bases, dictionary):
super().__init__(classname, bases, dictionary)
assert hasattr(cls, "load") and \
isinstance(getattr(cls, "load"),
collections.Callable), ("class '" +
classname + "' must provide a load() method")
assert hasattr(cls, "save") and \
isinstance(getattr(cls, "save"),
collections.Callable), ("class '" +
classname + "' must provide a save() method")
Classes that are to serve as metaclasses must inherit from the ultimate
metaclass base class, type, or one of its subclasses.
Note that this class is called when classes that use it are instantiated, in all
probability not very often, so the runtime cost is extremely low. Notice also
that we must perform the checks after the class has been created (using the
super() call), since only then will the class’s attributes be available in the class
itself. (The attributes are in the dictionary, but we prefer to work on the actual
initialized class when doing checks.)
In Python terminology, virtual does not mean the same thing as it does in C++ terminology.
383 ➤
Chapter 8. Advanced Programming Techniques
We could have checked that the load and save attributes are callable using
hasattr() to check that they have the __call__ attribute, but we prefer to
check whether they are instances of collections.Callable instead. The collections.Callable abstract base class provides the promise (but no guarantee) that
instances of its subclasses (or virtual subclasses) are callable.
Once the class has been created (using type.__new__() or a reimplementation
of __new__()), the metaclass is initialized by calling its __init__() method.
The arguments given to __init__() are cls, the class that’s just been created;
classname, the class’s name (also available from cls.__name__); bases, a list of
the class’s base classes (excluding object, and therefore possibly empty); and
dictionary that holds the attributes that became class attributes when the cls
class was created, unless we intervened in a reimplementation of the metaclass’s __new__() method.
Here are a couple of interactive examples that show what happens when we
create classes using the LoadableSaveable metaclass:
>>> class Bad(metaclass=Meta.LoadableSaveable):
def some_method(self): pass
Traceback (most recent call last):
AssertionError: class 'Bad' must provide a load() method
The metaclass specifies that classes using it must provide certain methods, and
when they don’t, as in this case, an AssertionError exception is raised.
>>> class Good(metaclass=Meta.LoadableSaveable):
def load(self): pass
def save(self): pass
>>> g = Good()
The Good class honors the metaclass’s API requirements, even if it doesn’t meet
our informal expectations of how it should behave.
We can also use metaclasses to change the classes that use them. If the change
involves the name, base classes, or dictionary of the class being created (e.g.,
its slots), then we need to reimplement the metaclass’s __new__() method; but
for other changes, such as adding methods or data attributes, reimplementing
__init__() is sufficient, although this can also be done in __new__(). We will now
look at a metaclass that modifies the classes it is used with purely through its
__new__() method.
As an alternative to using the @property and @name.setter decorators, we could
create classes where we use a simple naming convention to identify properties.
For example, if a class has methods of the form get_name() and set_name(),
we would expect the class to have a private __name property accessed using
Further Object-Oriented Programming
393 for getting and setting. This can all be done using a metaclass.
Here is an example of a class that uses this convention:
class Product(metaclass=AutoSlotProperties):
def __init__(self, barcode, description):
self.__barcode = barcode
self.description = description
def get_barcode(self):
return self.__barcode
def get_description(self):
return self.__description
def set_description(self, description):
if description is None or len(description) < 3:
self.__description = "<Invalid Description>"
self.__description = description
We must assign to the private __barcode property in the initializer since there
is no setter for it; another consequence of this is that barcode is a read-only
property. On the other hand, description is a readable/writable property. Here
are some examples of interactive use:
>>> product = Product("101110110", "8mm Stapler")
>>> product.barcode, product.description
('101110110', '8mm Stapler')
>>> product.description = "8mm Stapler (long)"
>>> product.barcode, product.description
('101110110', '8mm Stapler (long)')
If we attempt to assign to the bar code an AttributeError exception is raised
with the error text “can’t set attribute”.
If we look at the Product class’s attributes (e.g., using dir()), the only public
ones to be found are barcode and description. The get_name() and set_name()
methods are no longer there—they have been replaced with the name property.
And the variables holding the bar code and description are also private (__barcode and __description), and have been added as slots to minimize the class’s
memory use. This is all done by the AutoSlotProperties metaclass which is implemented in a single method:
class AutoSlotProperties(type):
def __new__(mcl, classname, bases, dictionary):
slots = list(dictionary.get("__slots__", []))
Chapter 8. Advanced Programming Techniques
for getter_name in [key for key in dictionary
if key.startswith("get_")]:
if isinstance(dictionary[getter_name],
name = getter_name[4:]
slots.append("__" + name)
getter = dictionary.pop(getter_name)
setter_name = "set_" + name
setter = dictionary.get(setter_name, None)
if (setter is not None and
isinstance(setter, collections.Callable)):
del dictionary[setter_name]
dictionary[name] = property(getter, setter)
dictionary["__slots__"] = tuple(slots)
return super().__new__(mcl, classname, bases, dictionary)
A metaclass’s __new__() class method is called with the metaclass, and the class
name, base classes, and dictionary of the class that is to be created. We must
use a reimplementation of __new__() rather than __init__() because we want
to change the dictionary before the class is created.
We begin by copying the __slots__ collection, creating an empty one if none
is present, and making sure we have a list rather than a tuple so that we can
modify it. For every attribute in the dictionary we pick out those that begin
with "get_" and that are callable, that is, those that are getter methods. For
each getter we add a private name to the slots to store the corresponding data;
for example, given getter get_name() we add __name to the slots. We then take a
reference to the getter and delete it from the dictionary under its original name
(this is done in one go using dict.pop()). We do the same for the setter if one is
present, and then we create a new dictionary item with the desired property
name as its key; for example, if the getter is get_name() the property name is
name. We set the item’s value to be a property with the getter and setter (which
might be None) that we have found and removed from the dictionary.
At the end we replace the original slots with the modified slots list which has
a private slot for each property that was added, and call on the base class to actually create the class, but using our modified dictionary. Note that in this case
we must pass the metaclass explicitly in the super() call; this is always the case
for calls to __new__() because it is a class method and not an instance method.
For this example we didn’t need to write an __init__() method because we have
done all the work in __new__(), but it is perfectly possible to reimplement both
__new__() and __init__() doing different work in each.
If we consider hand-cranked drills to be analogous to aggregation and inheritance and electric drills the analog of decorators and descriptors, then metaclasses are at the laser beam end of the scale when it comes to power and
Further Object-Oriented Programming
versatility. Metaclasses are the last tool to reach for rather than the first, except perhaps for application framework developers who need to provide powerful facilities to their users without making the users go through hoops to realize
the benefits on offer.
Functional-Style Programming
Functional-style programming is an approach to programming where computations are built up from combining functions that don’t modify their arguments and that don’t refer to or change the program’s state, and that provide
their results as return values. One strong appeal of this kind of programming
is that (in theory), it is much easier to develop functions in isolation and to debug functional programs. This is helped by the fact that functional programs
don’t have state changes, so it is possible to reason about their functions mathematically.
Three concepts that are strongly associated with functional programming are
mapping, filtering, and reducing. Mapping involves taking a function and an
iterable and producing a new iterable (or a list) where each item is the result
of calling the function on the corresponding item in the original iterable. This
is supported by the built-in map() function, for example:
list(map(lambda x: x ** 2, [1, 2, 3, 4]))
# returns: [1, 4, 9, 16]
The map() function takes a function and an iterable as its arguments and for
efficiency it returns an iterator rather than a list. Here we forced a list to be
created to make the result clearer:
[x ** 2 for x in [1, 2, 3, 4]]
# returns: [1, 4, 9, 16]
A generator expression can often be used in place of map(). Here we have used
a list comprehension to avoid the need to use list(); to make it a generator we
just have to change the outer brackets to parentheses.
Filtering involves taking a function and an iterable and producing a new iterable where each item is from the original iterable—providing the function
returns True when called on the item. The built-in filter() function supports this:
list(filter(lambda x: x > 0, [1, -2, 3, -4])) # returns: [1, 3]
The filter() function takes a function and an iterable as its arguments and
returns an iterator.
[x for x in [1, -2, 3, -4] if x > 0]
# returns: [1, 3]
Chapter 8. Advanced Programming Techniques
The filter() function can always be replaced with a generator expression or
with a list comprehension.
Reducing involves taking a function and an iterable and producing a single
result value. The way this works is that the function is called on the iterable’s
first two values, then on the computed result and the third value, then on the
computed result and the fourth value, and so on, until all the values have been
used. The functools module’s functools.reduce() function supports this. Here
are two lines of code that do the same computation:
functools.reduce(lambda x, y: x * y, [1, 2, 3, 4])
functools.reduce(operator.mul, [1, 2, 3, 4])
# returns: 24
# returns: 24
The operator module has functions for all of Python’s operators specifically to
make functional-style programming easier. Here, in the second line, we have
used the operator.mul() function rather than having to create a multiplication
function using lambda as we did in the first line.
Python also provides some built-in reducing functions: all(), which given an
iterable, returns True if all the iterable’s items return True when bool() is applied to them; any(), which returns True if any of the iterable’s items is True;
max(), which returns the largest item in the iterable; min(), which returns the
smallest item in the iterable; and sum(), which returns the sum of the iterable’s items.
Now that we have covered the key concepts, let us look at a few more examples.
We will start with a couple of ways to get the total size of all the files in list
functools.reduce(operator.add, (os.path.getsize(x) for x in files))
functools.reduce(operator.add, map(os.path.getsize, files))
Using map() is often shorter than the equivalent list comprehension or generator expression except where there is a condition. We’ve used operator.add() as
the addition function instead of lambda x, y: x + y.
If we only wanted to count the .py file sizes we can filter out non-Python files.
Here are three ways to do this:
functools.reduce(operator.add, map(os.path.getsize,
filter(lambda x: x.endswith(".py"), files)))
functools.reduce(operator.add, map(os.path.getsize,
(x for x in files if x.endswith(".py"))))
functools.reduce(operator.add, (os.path.getsize(x)
for x in files if x.endswith(".py")))
Arguably, the second and third versions are better because they don’t require
us to create a lambda function, but the choice between using generator
Functional-Style Programming
sions (or list comprehensions) and map() and filter() is most often purely a
matter of personal programming style.
Using map(), filter(), and functools.reduce() often leads to the elimination
of loops, as the examples we have seen illustrate. These functions are useful
when converting code written in a functional language, but in Python we
can usually replace map() with a list comprehension and filter() with a list
comprehension with a condition, and many cases of functools.reduce() can be
eliminated by using one of Python’s built-in functional functions such as all(),
any(), max(), min(), and sum(). For example:
sum(os.path.getsize(x) for x in files if x.endswith(".py"))
This achieves the same thing as the previous three examples, but is much
more compact.
369 ➤
In addition to providing functions for Python’s operators, the operator module
also provides the operator.attrgetter() and operator.itemgetter() functions,
the first of which we briefly met earlier in this chapter. Both of these return
functions which can then be called to extract the specified attributes or items.
Whereas slicing can be used to extract a sequence of part of a list, and slicing
with striding can be used to extract a sequence of parts (say, every third item
with L[::3]), operator.itemgetter() can be used to extract a sequence of arbitrary parts, for example, operator.itemgetter(4, 5, 6, 11, 18)(L). The function
returned by operator.itemgetter() does not have to be called immediately and
thrown away as we have done here; it could be kept and passed as the function
argument to map(), filter(), or functools.reduce(), or used in a dictionary, list,
or set comprehension.
When we want to sort we can specify a key function. This function can be any
function, for example, a lambda function, a built-in function or method (such
as str.lower()), or a function returned by operator.attrgetter(). For example,
assuming list L holds objects with a priority attribute, we can sort the list into
priority order like this: L.sort(key=operator.attrgetter("priority")).
In addition to the functools and operator modules already mentioned, the itertools module can also be useful for functional-style programming. For example, although it is possible to iterate over two or more lists by concatenating
them, an alternative is to use itertools.chain() like this:
for value in itertools.chain(data_list1, data_list2, data_list3):
total += value
The itertools.chain() function returns an iterator that gives successive values
from the first sequence it is given, then successive values from the second
sequence, and so on until all the values from all the sequences are used. The
itertools module has many other functions, and its documentation gives many
Chapter 8. Advanced Programming Techniques
small yet useful examples and is well worth reading. (Note also that a couple
of new functions were added to the itertools module with Python 3.1.)
Partial Function Application
Partial function application is the creation of a function from an existing
function and some arguments to produce a new function that does what the
original function did, but with some arguments fixed so that callers don’t have
to pass them. Here’s a very simple example:
enumerate1 = functools.partial(enumerate, start=1)
for lino, line in enumerate1(lines):
process_line(i, line)
The first line creates a new function, enumerate1(), that wraps the given function (enumerate()) and a keyword argument (start=1) so that when enumerate1()
is called it calls the original function with the fixed argument—and with any
other arguments that are given at the time it is called, in this case lines. Here
we have used the enumerate1() function to provide conventional line counting
starting from line 1.
Using partial function application can simplify our code, especially when we
want to call the same functions with the same arguments again and again. For
example, instead of specifying the mode and encoding arguments every time
we call open() to process UTF-8 encoded text files, we could create a couple of
functions with these arguments fixed:
reader = functools.partial(open, mode="rt", encoding="utf8")
writer = functools.partial(open, mode="wt", encoding="utf8")
Now we can open text files for reading by calling reader(filename) and for
writing by calling writer(filename).
One very common use case for partial function application is in GUI (Graphical
User Interface) programming (covered in Chapter 15), where it is often convenient to have one particular function called when any one of a set of buttons is
pressed. For example:
loadButton = tkinter.Button(frame, text="Load",
command=functools.partial(doAction, "load"))
saveButton = tkinter.Button(frame, text="Save",
command=functools.partial(doAction, "save"))
This example uses the tkinter GUI library that comes as standard with
Python. The tkinter.Button class is used for buttons—here we have created
two, both contained inside the same frame, and each with a text that indicates
its purpose. Each button’s command argument is set to the function that tkinter
Functional-Style Programming
must call when the button is pressed, in this case the doAction() function. We
have used partial function application to ensure that the first argument given
to the doAction() function is a string that indicates which button called it so
that doAction() is able to decide what action to perform.
Coroutines are functions whose processing can be suspended and resumed at
specific points. So, typically, a coroutine will execute up to a certain statement,
then suspend execution while waiting for some data. At this point other parts
of the program can continue to execute (usually other coroutines that aren’t
suspended). Once the data is received the coroutine resumes from the point it
was suspended, performs processing (presumably based on the data it got), and
possibly sending its results to another coroutine. Coroutines are said to have
multiple entry and exit points, since they can have more than one place where
they suspend and resume.
Coroutines are useful when we want to apply multiple functions to the same
pieces of data, or when we want to create data processing pipelines, or when
we want to have a master function with slave functions. Coroutines can also
be used to provide simpler and lower-overhead alternatives to threading. A
few coroutine-based packages that provide lightweight threading are available
from the Python Package Index,
341 ➤
In Python, a coroutine is a function that takes its input from a yield expression.
It may also send results to a receiver function (which itself must be a coroutine). Whenever a coroutine reaches a yield expression it suspends waiting for
data; and once it receives data, it resumes execution from that point. A coroutine can have more than one yield expression, although each of the coroutine
examples we will review has only one.
Performing Independent Actions on Data
If we want to perform a set of independent operations on some data, the
conventional approach is to apply each operation in turn. The disadvantage of
this is that if one of the operations is slow, the program as a whole must wait
for the operation to complete before going on to the next one. A solution to this
is to use coroutines. We can implement each operation as a coroutine and then
start them all off. If one is slow it won’t affect the others—at least not until
they run out of data to process—since they all operate independently.
Figure 8.2 illustrates the use of coroutines for concurrent processing. In the figure, three coroutines (each presumably doing a different job) process the same
two data items—and take different amounts of time to do their work. In the
figure, coroutine1() works quite quickly, coroutine2() works slowly, and coroutine3() varies. Once all three coroutines have been given their initial data
Chapter 8. Advanced Programming Techniques
Create coroutines
Process "a"
Process "a"
Process "a"
Process "a"
Process "a"
Process "b"
Process "a"
Process "a"
Process "b"
Process "a"
("b" pending)
Process "a"
Process "a"
("b" pending)
Process "b"
Process "b"
Process "b"
Process "b"
Process "b"
Figure 8.2 Sending two items of data to three coroutines
to process, if one is ever waiting (because it finishes first), the others continue
to work, which minimizes processor idle time. Once we are finished using the
coroutines we call close() on each of them; this stops them from waiting for
more data, which means they won’t consume any more processor time.
To create a coroutine in Python, we simply create a function that has at
least one yield expression—normally inside an infinite loop. When a yield is
reached the coroutine’s execution is suspended waiting for data. Once the data
is received the coroutine resumes processing (from the yield expression onward), and when it has finished it loops back to the yield to wait for more data.
While one or more coroutines are suspended waiting for data, another one can
execute. This can produce greater throughput than simply executing functions
one after the other linearly.
We will show how performing independent operations works in practice by
applying several regular expressions to the text in a set of HTML files. The
purpose is to output each file’s URLs and level 1 and level 2 headings. We’ll
start by looking at the regular expressions, then the creation of the coroutine
“matchers”, and then we will look at the coroutines and how they are used.
URL_RE = re.compile(r"""href=(?P<quote>['"])(?P<url>[^\1]+?)"""
r"""(?P=quote)""", re.IGNORECASE)
H1_RE = re.compile(r"<h1>(?P<h1>.+?)</h1>", flags)
H2_RE = re.compile(r"<h2>(?P<h2>.+?)</h2>", flags)
Functional-Style Programming
These regular expressions (“regexes” from now on) match an HTML href’s URL
and the text contained in <h1> and <h2> header tags. (Regular expressions are
covered in Chapter 13; understanding them is not essential to understanding
this example.)
receiver = reporter()
matchers = (regex_matcher(receiver, URL_RE),
regex_matcher(receiver, H1_RE),
regex_matcher(receiver, H2_RE))
341 ➤
Since coroutines always have a yield expression, they are generators. So
although here we create a tuple of matcher coroutines, in effect we are creating
a tuple of generators. Each regex_matcher() is a coroutine that takes a receiver
function (itself a coroutine) and a regex to match. Whenever the matcher
matches it sends the match to the receiver.
def regex_matcher(receiver, regex):
while True:
text = (yield)
for match in regex.finditer(text):
The matcher starts by entering an infinite loop and immediately suspends
execution waiting for the yield expression to return a text to apply the regex
to. Once the text is received, the matcher iterates over every match it makes,
sending each one to the receiver. Once the matching has finished the coroutine
loops back to the yield and again suspends waiting for more text.
There is one tiny problem with the (undecorated) matcher—when it is first
created it should commence execution so that it advances to the yield ready to
receive its first text. We could do this by calling the built-in next() function on
each coroutine we create before sending it any data. But for convenience we
have created the @coroutine decorator to do this for us.
def coroutine(function):
def wrapper(*args, **kwargs):
generator = function(*args, **kwargs)
return generator
return wrapper
356 ➤
The @coroutine decorator takes a coroutine function, and calls the built-in
next() function on it—this causes the function to be executed up to the first
yield expression, ready to receive data.
Chapter 8. Advanced Programming Techniques
Now that we have seen the matcher coroutine we will look at how the matchers
are used, and then we will look at the reporter() coroutine that receives the
matchers’ outputs.
for file in sys.argv[1:]:
html = open(file, encoding="utf8").read()
for matcher in matchers:
for matcher in matchers:
The program reads the filenames listed on the command line, and for each one
prints the filename and then reads the file’s entire text into the html variable
using the UTF-8 encoding. Then the program iterates over all the matchers
(three in this case), and sends the text to each of them. Each matcher then
proceeds independently, sending each match it makes to the reporter coroutine.
At the end we call close() on each matcher and on the reporter—this terminates them, since otherwise they would continue (suspended) waiting for text
(or matches in the case of the reporter) since they contain infinite loops.
def reporter():
ignore = frozenset({"style.css", "favicon.png", "index.html"})
while True:
match = (yield)
if match is not None:
groups = match.groupdict()
if "url" in groups and groups["url"] not in ignore:
URL:", groups["url"])
elif "h1" in groups:
H1: ", groups["h1"])
elif "h2" in groups:
H2: ", groups["h2"])
The reporter() coroutine is used to output results. It was created by the statement receiver = reporter() which we saw earlier, and passed as the receiver
argument to each of the matchers. The reporter() waits (is suspended) until
a match is sent to it, then it prints the match’s details, and then it waits again,
in an endless loop—stopping only if close() is called on it.
Using coroutines like this may produce performance benefits, but does require
us to adopt a somewhat different way of thinking about processing.
Functional-Style Programming
Composing Pipelines
Sometimes it is useful to create data processing pipelines. A pipeline is simply
the composition of one or more functions where data items are sent to the first
function, which then either discards the item (filters it out) or passes it on to the
next function (either as is or transformed in some way). The second function
receives the item from the first function and repeats the process, discarding
or passing on the item (possibly transformed in a different way) to the next
function, and so on. Items that reach the end are then output in some way.
395 ➤
Pipelines typically have several components, one that acquires data, one or
more that filter or transform data, and one that outputs results. This is exactly
the functional-style approach to programming that we discussed earlier in the
section when we looked at composing some of Python’s built-in functions, such
as filter() and map().
One benefit of using pipelines is that we can read data items incrementally,
often one at a time, and have to give the pipeline only enough data items to
fill it (usually one or a few items per component). This can lead to significant
memory savings compared with, say, reading an entire data set into memory
and then processing it all in one go.
pipeline = get_data(
Read "a"
Read "b"
Process "a"
Read "c"
Process "b"
Output "a"
Read "d"
Process "c"
Output "b"
Read "e"
Drop "d"
Output "c"
Read "f"
Process "e"
Process "f"
Output "e"
Output "f"
Close coroutines
Figure 8.3 A three-step coroutine pipeline processing six items of data
Figure 8.3 illustrates a simple three component pipeline. The first component
of the pipeline (get_data()) acquires each data item to be processed in turn.
The second component (process()) processes the data—and may drop unwanted
data items—there could be any number of other processing/filtering components, of course. The last component (reporter()) outputs results. In the figure,
Chapter 8. Advanced Programming Techniques
items "a", "b", "c", "e", and "f" are processed and produce output, while item
"d" is dropped.
The pipeline shown in Figure 8.3 is a filter, since each data item is passed
through unchanged and is either dropped or output in its original form. The
end points of pipelines tend to perform the same roles: acquiring data items
and outputting results. But between these we can have as many components
as necessary, each filtering or transforming or both. And in some cases, composing the components in different orders can produce pipelines that do different things.
We will start out by looking at a theoretical example to get a better idea of how
coroutine-based pipelines work, and then we will look at a real example.
Suppose we have a sequence of floating-point numbers and we want to process
them in a multicomponent pipeline such that we transform each number into
an integer (by rounding), but drop any numbers that are out of range (< 0 or >=
10). If we had the four coroutine components, acquire() (get a number), to_int()
(transform a number by rounding and converting to an integer), check() (pass
on a number that is in range; drop a number that is out of range), and output()
(output a number), we could create the pipeline like this:
pipe = acquire(to_int(check(output())))
We would then send numbers into the pipeline by calling pipe.send(). We’ll
look at the progress of the numbers 4.3 and 9.6 as they go through the pipeline,
using a different visualization from the step-by-step figures used earlier:
pipe.send(4.3) → acquire(4.3) → to_int(4.3) → check(4) → output(4)
pipe.send(9.6) → acquire(9.6) → to_int(9.6) → check(10)
Notice that for 9.6 there is no output. This is because the check() coroutine
received 10, which is out of range (>= 10), and so it was filtered out.
Let’s see what would happen if we created a different pipeline, but using the
same components:
pipe = acquire(check(to_int(output())))
This simply performs the filtering (check()) before the transforming (to_int()).
Here is how it would work for 4.3 and 9.6:
pipe.send(4.3) → acquire(4.3) → check(4.3) → to_int(4.3) → output(4)
pipe.send(9.6) → acquire(9.6) → check(9.6) → to_int(9.6) → output(10)
Here we have incorrectly output 10, even though it is out of range. This is
because we applied the check() component first, and since this received an
in-range value of 9.6, it simply passed it on. But the to_int() component
rounds the numbers it gets.
Functional-Style Programming
We will now review a concrete example—a file matcher that reads all the
filenames given on the command line (including those in the directories given
on the command line, recursively), and that outputs the absolute paths of those
files that meet certain criteria.
We will start by looking at how pipelines are composed, and then we will
look at the coroutines that provide the pipeline components. Here is the simplest pipeline:
pipeline = get_files(receiver)
224 ➤
This pipeline prints every file it is given (or all the files in the directory it
is given, recursively). The get_files() function is a coroutine that yields the
filenames and the receiver is a reporter() coroutine—created by receiver =
reporter()—that simply prints each filename it receives. This pipeline does
little more than the os.walk() function (and in fact uses that function), but we
can use its components to compose more sophisticated pipelines.
pipeline = get_files(suffix_matcher(receiver, (".htm", ".html")))
This pipeline is created by composing the get_files() coroutine together with
the suffix_matcher() coroutine. It prints only HTML files.
Coroutines composed like this can quickly become difficult to read, but there
is nothing to stop us from composing a pipeline in stages—although for this
approach we must create the components in last-to-first order.
pipeline = size_matcher(receiver, minimum=1024 ** 2)
pipeline = suffix_matcher(pipeline, (".png", ".jpg", ".jpeg"))
pipeline = get_files(pipeline)
This pipeline only matches files that are at least one megabyte in size, and that
have a suffix indicating that they are images.
How are these pipelines used? We simply feed them filenames or paths and
they take care of the rest themselves.
for arg in sys.argv[1:]:
Notice that it doesn’t matter which pipeline we are using—it could be the
one that prints all the files, or the one that prints HTML files, or the images
one—they all work in the same way. And in this case, all three of the pipelines
are filters—any filename they get is either passed on as is to the next component (and in the case of the reporter(), printed), or dropped because they don’t
meet the criteria.
Before looking at the get_files() and the matcher coroutines, we will look at
the trivial reporter() coroutine (passed as receiver) that outputs the results.
Chapter 8. Advanced Programming Techniques
def reporter():
while True:
filename = (yield)
@coroutine dec-
401 ➤
224 ➤
We have used the same @coroutine decorator that we created in the previous
The get_files() coroutine is essentially a wrapper around the os.walk()
function and that expects to be given paths or filenames to work on.
def get_files(receiver):
while True:
path = (yield)
if os.path.isfile(path):
for root, dirs, files in os.walk(path):
for filename in files:
os.path.join(root, filename)))
This coroutine has the now-familiar structure: an infinite loop in which we wait
for the yield to return a value that we can process, and then we send the result
to the receiver.
def suffix_matcher(receiver, suffixes):
while True:
filename = (yield)
if filename.endswith(suffixes):
This coroutine looks simple—and it is—but notice that it sends only filenames that match the suffixes, so any that don’t match are filtered out of
the pipeline.
def size_matcher(receiver, minimum=None, maximum=None):
while True:
filename = (yield)
size = os.path.getsize(filename)
if ((minimum is None or size >= minimum) and
(maximum is None or size <= maximum)):
Functional-Style Programming
This coroutine is almost identical to suffix_matcher(), except that it filters out
files whose size is not in the required range, rather than those which don’t have
a matching suffix.
The pipeline we have created suffers from a couple of problems. One problem
is that we never close any of the coroutines. In this case it doesn’t matter,
since the program terminates once the processing is finished, but it is probably
better to get into the habit of closing coroutines when we are finished with
them. Another problem is that potentially we could be asking the operating
system (under the hood) for different pieces of information about the same file
in several parts of the pipeline—and this could be slow. A solution is to modify
the get_files() coroutine so that it returns (filename, os.stat()) 2-tuples for
each file rather than just filenames, and then pass these 2-tuples through the
pipeline.★ This would mean that we acquire all the relevant information just
once per file. You’ll get the chance to solve both of these problems, and to add
additional functionality, in an exercise at the end of the chapter.
Creating coroutines for use in pipelines requires a certain reorientation of
thinking. However, it can pay off handsomely in terms of flexibility, and for
large data sets can help minimize the amount of data held in memory as well
as potentially resulting in faster throughput.
372 ➤
378 ➤
In this section we combine descriptors with class decorators to create a
powerful mechanism for creating validated attributes.
Up to now if we wanted to ensure that an attribute was set to only a valid value
we have relied on properties (or used getter and setter methods). The disadvantage of such approaches is that we must add validating code for every attribute
in every class that needs it. What would be much more convenient and easier to
maintain, is if we could add attributes to classes with the necessary validation
built in. Here is an example of the syntax we would like to use:
@valid_string("name", empty_allowed=False)
@valid_string("productid", empty_allowed=False,
@valid_string("category", empty_allowed=False, acceptable=
frozenset(["Consumables", "Hardware", "Software", "Media"]))
@valid_number("price", minimum=0, maximum=1e6)
@valid_number("quantity", minimum=1, maximum=1000)
class StockItem:
The os.stat() function takes a filename and returns a named tuple with various items of
information about the file, including its size, mode, and last modified date/time.
Chapter 8. Advanced Programming Techniques
def __init__(self, name, productid, category, price, quantity): = name
self.productid = productid
self.category = category
self.price = price
self.quantity = quantity
The StockItem class’s attributes are all validated. For example, the productid
attribute can be set only to a nonempty string that starts with three uppercase
letters and ends with four digits, the category attribute can be set only to a
nonempty string that is one of the specified values, and the quantity attribute
can be set only to a number between 1 and 1 000 inclusive. If we try to set an
invalid value an exception is raised.
378 ➤
The validation is achieved by combining class decorators with descriptors. As
we noted earlier, class decorators can take only a single argument—the class
they are to decorate. So here we have used the technique shown when we first
discussed class decorators, and have the valid_string() and valid_number()
functions take whatever arguments we want, and then return a decorator,
which in turn takes the class and returns a modified version of the class.
Let’s now look at the valid_string() function:
def valid_string(attr_name, empty_allowed=True, regex=None,
def decorator(cls):
name = "__" + attr_name
def getter(self):
return getattr(self, name)
def setter(self, value):
assert isinstance(value, str), (attr_name +
" must be a string")
if not empty_allowed and not value:
raise ValueError("{0} may not be empty".format(
if ((acceptable is not None and value not in acceptable) or
(regex is not None and not regex.match(value))):
raise ValueError("{attr_name} cannot be set to "
setattr(self, name, value)
setattr(cls, attr_name, GenericDescriptor(getter, setter))
return cls
return decorator
The function starts by creating a class decorator function which takes a class as
its sole argument. The decorator adds two attributes to the class it decorates: a
private data attribute and a descriptor. For example, when the valid_string()
➤ 489
function is called with the name “productid”, the StockItem class gains the
attribute __productid which holds the product ID’s value, and the descriptor productid attribute which is used to access the value. For example, if we
create an item using item = StockItem("TV", "TVA4312", "Electrical", 500, 1),
we can get the product ID using item.productid and set it using, for example,
item.productid = "TVB2100".
The getter function created by the decorator simply uses the global getattr()
function to return the value of the private data attribute. The setter function
incorporates the validation, and at the end, uses setattr() to set the private
data attribute to the new (and valid) value. In fact, the private data attribute
is only created the first time it is set.
Once the getter and setter functions have been created we use setattr() once
again, this time to create a new class attribute with the given name (e.g.,
productid), and with its value set to be a descriptor of type GenericDescriptor. At the end, the decorator function returns the modified class, and the
valid_string() function returns the decorator function.
The valid_number() function is structurally identical to the valid_string()
function, only differing in the arguments it accepts and in the validation code
in the setter, so we won’t show it here. (The complete source code is in the module.)
The last thing we need to cover is the GenericDescriptor, and that turns out to
be the easiest part:
class GenericDescriptor:
def __init__(self, getter, setter):
self.getter = getter
self.setter = setter
def __get__(self, instance, owner=None):
if instance is None:
return self
return self.getter(instance)
def __set__(self, instance, value):
return self.setter(instance, value)
The descriptor is used to hold the getter and setter functions for each attribute
and simply passes on the work of getting and setting to those functions.
Chapter 8. Advanced Programming Techniques
In this chapter we learned a lot more about Python’s support for procedural
and object-oriented programming, and got a taste of Python’s support for
functional-style programming.
In the first section we learned how to create generator expressions, and covered
generator functions in more depth. We also learned how to dynamically import
modules and how to access functionality from such modules, as well as how to
dynamically execute code. In this section we saw examples of how to create
and use recursive functions and nonlocal variables. We also learned how to
create custom function and method decorators, and how to write and make use
of function annotations.
In the chapter’s second section we studied a variety of different and more advanced aspects of object-oriented programming. First we learned more about
attribute access, for example, using the __getattr__() special method. Then
we learned about functors and saw how we could use them to provide functions
with state—something that can also be achieved by adding properties to functions or using closures, both covered in this chapter. We learned how to use
the with statement with context managers and how to create custom context
managers. Since Python’s file objects are also context managers, from now on
we will do our file handling using try with … except structures that ensure that
opened files are closed without the need for finally blocks.
The second section continued with coverage of more advanced object-oriented
features, starting with descriptors. These can be used in a wide variety of ways
and are the technology that underlies many of Python’s standard decorators
such as @property and @classmethod. We learned how to create custom descriptors and saw three very different examples of their use. Next we studied class
decorators and saw how we could modify a class in much the same way that a
function decorator can modify a function.
In the last three subsections of the second section we learned about Python’s
support for ABCs (abstract base classes), multiple inheritance, and metaclasses. We learned how to make our own classes fit in with Python’s standard ABCs
and how to create our own ABCs. We also saw how to use multiple inheritance
to unify the features of different classes together in a single class. And from the
coverage of metaclasses we learned how to intervene when a class (as opposed
to an instance of a class) is created and initialized.
The penultimate section introduced some of the functions and modules that
Python provides to support functional-style programming. We learned how to
use the common functional idioms of mapping, filtering, and reducing. We also
learned how to create partial functions and how to create and use coroutines.
And the last section showed how to combine class decorators with descriptors to
provide a powerful and flexible mechanism for creating validated attributes.
This chapter completes our coverage of the Python language itself. Not every
feature of the language has been covered here and in the previous chapters,
but those that have not are obscure and rarely used. None of the subsequent
chapters introduces new language features, although all of them make use
of modules from the standard library that have not been covered before, and
some of them take techniques shown in this and earlier chapters further
than we have seen so far. Furthermore, the programs shown in the following
chapters have none of the constraints that have applied previously (i.e., to only
use aspects of the language that had been covered up to the point they were
introduced), so they are the book’s most idiomatic examples.
None of the first three exercises described here requires writing a lot of code—
although the fourth one does—and none of them are easy!
1. Copy the program and delete its get_function() functions,
and all but one of its load_modules() functions. Add a GetFunction functor
class that has two caches, one to hold functions that have been found and
one to hold functions that could not be found (to avoid repeatedly looking
for a function in a module that does not have the function). The only modifications to main() are to add get_function = GetFunction() before the loop,
and to use a with statement to avoid the need for a finally block. Also,
check that the module functions are callable using collections.Callable
rather than using hasattr(). The class can be written in about twenty lines.
A solution is in
2. Create a new module file and in it define three functions: is_ascii() that
returns True if all the characters in the given string have code points less
than 127; is_ascii_punctuation() that returns True if all the characters
are in the string.punctuation string; and is_ascii_printable() that returns
True if all the characters are in the string.printable string. The last two
are structurally the same. Each function should be created using lambda
and can be done in one or two lines using functional-style code. Be sure to
add a docstring for each one with doctests and to make the module run the
doctests. The functions require only three to five lines for all three of them,
with the whole module fewer than 25 lines including doctests. A solution
is given in
3. Create a new module file and in it define the Atomic context manager class.
This class should work like the AtomicList class shown in this chapter, except that instead of working only with lists it should work with any mutable collection type. The __init__() method should check the suitability
Chapter 8. Advanced Programming Techniques
of the container, and instead of storing a shallow/deep copy flag it should
assign a suitable function to the self.copy attribute depending on the
flag and call the copy function in the __enter__() method. The __exit__()
method is slightly more involved because replacing the contents of lists
is different than for sets and dictionaries—and we cannot use assignment
because that would not affect the original container. The class itself can
be written in about thirty lines, although you should also include doctests.
A solution is given in which is about one hundred fifty lines including doctests.
4. Create a program that finds files based on specified criteria (rather like the
Unix find program). The usage should be options files_or_paths.
All the options are optional, and without them all the files listed on the
command line and all the files in the directories listed on the command
line (and in their directories, recursively) should be listed. The options
should restrict which files are output as follows: -d or --days integer discards any files older than the specified number of days; -b or --bigger integer discards any files smaller than the specified number of bytes; -s or
--smaller integer discards any files bigger than the specified number of
bytes; -o or --output what where what is “date”, “size”, or “date,size” (either
way around) specifies what should be output—filenames should always be
output; -u or --suffix discards any files that don’t have a matching suffix.
(Multiple suffixes can be given if comma-separated.) For both the bigger
and smaller options, if the integer is followed by “k” it should be treated as
kilobytes and multipled by 1024, and similarly if followed by “m” treated
as megabytes and multiplied by 10242.
For example, -d1 -o date,size *.* will find all files modified today
(strictly, the past 24 hours), and output their name, date, and size. Similarly, -b1m -u png,jpg,jpeg -o size *.* will find all image files bigger
than one megabyte and output their names and sizes.
Implement the program’s logic by creating a pipeline using coroutines to
provide matchers, similar to what we saw in the coroutines subsection,
only this time pass (filename, os.stat()) 2-tuples for each file rather than
just filenames. Also, try to close all the pipeline components at the end. In
the solution provided, the biggest single function is the one that handles
the command-line options. The rest is fairly straightforward, but not
trivial. The solution is around 170 lines.
● Debugging
● Unit Testing
● Profiling
Debugging, Testing, and
Writing programs is a mixture of art, craft, and science, and because it is done
by humans, mistakes are made. Fortunately, there are techniques we can use
to help avoid problems in the first place, and techniques for identifying and
fixing mistakes when they become apparent.
Mistakes fall into several categories. The quickest to reveal themselves and
the easiest to fix are syntax errors, since these are usually due to typos. More
challenging are logical errors—with these, the program runs, but some aspect
of its behavior is not what we intended or expected. Many errors of this kind
can be prevented from happening by using TDD (Test Driven Development),
where when we want to add a new feature, we begin by writing a test for the
feature—which will fail since we haven’t added the feature yet—and then implement the feature itself. Another mistake is to create a program that has
needlessly poor performance. This is almost always due to a poor choice of algorithm or data structure or both. However, before attempting any optimization we should start by finding out exactly where the performance bottleneck
lies—since it might not be where we expect—and then we should carefully decide what optimization we want to do, rather than working at random.
In this chapter’s first section we will look at Python’s tracebacks to see how to
spot and fix syntax errors and how to deal with unhandled exceptions. Then
we will see how to apply the scientific method to debugging to make finding
errors as fast and painless as possible. We will also look at Python’s debugging
support. In the second section we will look at Python’s support for writing unit
tests, and in particular the doctest module we saw earlier (in Chapter 5 and
Chapter 6), and the unittest module. We will see how to use these modules to
support TDD. In the chapter’s final section we will briefly look at profiling, to
identify performance hot spots so that we can properly target our optimization
Chapter 9. Debugging, Testing, and Profiling
In this section we will begin by looking at what Python does when there is a
syntax error, then at the tracebacks that Python produces when unhandled exceptions occur, and then we will see how to apply the scientific method to debugging. But before all that we will briefly discuss backups and version control.
When editing a program to fix a bug there is always the risk that we end up
with a program that has the original bug plus new bugs, that is, it is even worse
than it was when we started! And if we haven’t got any backups (or we have
but they are several changes out of date), and we don’t use version control, it
could be very hard to even get back to where we just had the original bug.
Making regular backups is an essential part of programming—no matter
how reliable our machine and operating system are and how rare failures
are—since failures still occur. But backups tend to be coarse-grained, with files
hours or even days old.
Version control systems allow us to incrementally save changes at whatever
level of granularity we want—every single change, or every set of related
changes, or simply every so many minutes’ worth of work. Version control
systems allow us to apply changes (e.g., to experiment with bugfixes), and if
they don’t work out, we can revert the changes back to the last “good” version
of the code. So before starting to debug, it is always best to check our code into
the version control system so that we have a known position that we can revert
to if we get into a mess.
There are many good cross-platform open source version control systems
available—this book uses Bazaar (, but other popular ones
include Mercurial (, Git (, and Subversion
( Incidentally, both Bazaar and Mercurial are mostly
written in Python. None of these systems is hard to use (at least for the basics),
but using any one of them will help avoid a lot of unnecessary pain.
Dealing with Syntax Errors
If we try to run a program that has a syntax error, Python will stop execution
and print the filename, line number, and offending line, with a caret (^) underneath indicating exactly where the error was detected. Here’s an example:
File "", line 383
if BlockOutput.save_blocks_as_svg(blocks, svg)
SyntaxError: invalid syntax
Did you see the error? We’ve forgotten to put a colon at the end of the if
statement’s condition.
Here is an example that comes up quite often, but where the problem isn’t at
all obvious:
File "", line 385
except ValueError as err:
SyntaxError: invalid syntax
There is no syntax error in the line indicated, so both the line number and the
caret’s position are wrong. In general, when we are faced with an error that
we are convinced is not in the specified line, in almost every case the error will
be in an earlier line. Here’s the code from the try to the except where Python
is reporting the error to be—see if you can spot the error before reading the
explanation that follows the code:
blocks = parse(blocks)
svg = file.replace(".blk", ".svg")
if not BlockOutput.save_blocks_as_svg(blocks, svg):
print("Error: failed to save {0}".format(svg)
except ValueError as err:
Did you spot the problem? It is certainly easy to miss since it is on the line
before the one that Python reports as having the error. We have closed the
str.format() method’s parentheses, but not the print() function’s parentheses,
that is, we are missing a closing parenthesis at the end of the line, but Python
didn’t realize this until it reached the except keyword on the following line.
Missing the last parenthesis on a line is quite common, especially when using
print() with str.format(), but the error is usually reported on the following
line. Similarly, if a list’s closing bracket, or a set or dictionary’s closing brace
is missing, Python will normally report the problem as being on the next (nonblank) line. On the plus side, syntax errors like these are trivial to fix.
Dealing with Runtime Errors
If an unhandled exception occurs at runtime, Python will stop executing our
program and print a traceback. Here is an example of a traceback for an
unhandled exception:
Traceback (most recent call last):
File "", line 392, in <module>
File "", line 381, in main
Chapter 9. Debugging, Testing, and Profiling
blocks = parse(blocks)
File "", line 174, in recursive_descent_parse
return data.stack[1]
IndexError: list index out of range
Tracebacks (also called backtraces) like this should be read from their last line
back toward their first line. The last line specifies the unhandled exception
that occurred. Above this line, the filename, line number, and function name,
followed by the line that caused the exception, are shown (spread over two
lines). If the function where the exception was raised was called by another
function, that function’s filename, line number, function name, and calling
line are shown above. And if that function was called by another function the
same applies, all the way up to the beginning of the call stack. (Note that the
filenames in tracebacks are given with their path, but in most cases we have
omitted paths from the examples for the sake of clarity.)
Function references
340 ➤
So in this example, an IndexError occurred, meaning that data.stack is some
kind of sequence, but has no item at position 1. The error occurred at line
174 in the program’s recursive_descent_parse() function, and that
function was called at line 381 in the main() function. (The reason that the
function’s name is different at line 381, that is, parse() instead of recursive_descent_parse(), is that the parse variable is set to one of several different
functions depending on the command-line arguments given to the program; in
the common case the names always match.) The call to main() was made at line
392, and this is the statement at which program execution commenced.
Although at first sight the traceback looks intimidating, now that we understand its structure it is easy to see how useful it is. In this case it tells us exactly where to look for the problem, although of course we must work out for
ourselves what the solution is.
Here is another example traceback:
Traceback (most recent call last):
File "", line 392, in <module>
File "", line 383, in main
if BlockOutput.save_blocks_as_svg(blocks, svg):
File "", line 141, in save_blocks_as_svg
widths, rows = compute_widths_and_rows(cells, SCALE_BY)
File "", line 95, in compute_widths_and_rows
width = len(cell.text) // cell.columns
ZeroDivisionError: integer division or modulo by zero
Here, the problem has occurred in a module ( that is called
by the program. This traceback leads us to where the problem
became apparent, but not to where it occurred. The value of cell.columns is
clearly 0 in the module’s compute_widths_and_rows() function
on line 95—after all, that is what caused the ZeroDivisionError exception to
be raised—but we must look at the preceding lines to find where and why
cell.columns was given this incorrect value.
In some cases the traceback reveals an exception that occurred in Python’s
standard library or in a third-party library. Although this could mean a bug in
the library, in almost every case it is due to a bug in our own code. Here is an
example of such a traceback, using Python 3.0:
Traceback (most recent call last):
File "", line 392, in <module>
File "", line 379, in main
blocks = open(file, encoding="utf8").read()
File "/usr/lib/python3.0/lib/python3.0/", line 278, in __new__
return open(*args, **kwargs)
File "/usr/lib/python3.0/lib/python3.0/", line 222, in open
File "/usr/lib/python3.0/lib/python3.0/", line 619, in __init__
_fileio._FileIO.__init__(self, name, mode, closefd)
IOError: [Errno 2] No such file or directory: 'hierarchy.blk'
The IOError exception at the end tells us clearly what the problem is. But
the exception was raised in the standard library’s io module. In such cases
it is best to keep reading upward until we find the first file listed that is our
program’s file (or one of the modules we have created for it). So in this case we
find that the first reference to our program is to file, line 379, in the
main() function. It looks like we have a call to open() but have not put the call
inside a try … except block or used a with statement.
Python 3.1 is a bit smarter than Python 3.0 and realizes that we want to find
the mistake in our own code, not in the standard library, so it produces a much
more compact and helpful traceback. For example:
Traceback (most recent call last):
File "", line 392, in <module>
File "", line 379, in main
blocks = open(file, encoding="utf8").read()
IOError: [Errno 2] No such file or directory: 'hierarchy.blk'
This eliminates all the irrelevant detail and makes it easy to see what the
problem is (on the bottom line) and where it occurred (the lines above it).
So no matter how big the traceback is, the last line always specifies the unhandled exception, and we just have to work back until we find our program’s file
Chapter 9. Debugging, Testing, and Profiling
or one of our own modules listed. The problem will almost certainly be on the
line Python specifies, or on an earlier line.
This particular example illustrates that we should modify the program to cope gracefully when given the names of nonexistent files. This is a
usability error, and it should also be classified as a logical error, since terminating and printing a traceback cannot be considered to be acceptable program
In fact, as a matter of good policy and courtesy to our users, we should always
catch all relevant exceptions, identifying the specific ones that we consider to be
possible, such as EnvironmentError. In general, we should not use the catchalls
of except: or except Exception:, although using the latter at the top level of our
program to avoid crashes might be appropriate—but only if we always report
any exceptions it catches so that they don’t go silently unnoticed.
Exceptions that we catch and cannot recover from should be reported in the
form of error messages, rather than exposing our users to tracebacks which
look scary to the uninitiated. For GUI programs the same applies, except that
normally we would use a message box to notify the user of a problem. And
for server programs that normally run unattended, we should write the error
message to the server’s log.
Python’s exception hierarchy was designed so that catching Exception doesn’t
quite cover all the exceptions. In particular, it does not catch the KeyboardInterrupt exception, so for console applications if the user presses Ctrl+C, the program
will terminate. If we choose to catch this exception, there is a risk that we could
lock the user into a program that they cannot terminate. This arises because
a bug in our exception handling code might prevent the program from terminating or the exception propagating. (Of course, even an “uninterruptible”
program can have its process killed, but not all users know how.) So if we do
catch the KeyboardInterrupt exception we must be extremely careful to do the
minimum amount of saving and clean up that is necessary—and then terminate the program. And for programs that don’t need to save or clean up, it is
best not to catch KeyboardInterrupt at all, and just let the program terminate.
One of Python 3’s great virtues is that it makes a clear distinction between raw
bytes and strings. However, this can sometimes lead to unexpected exceptions
occurring when we pass a bytes object where a str is expected or vice versa.
For example:
Traceback (most recent call last):
File "", line 918, in <module>
print(datetime.datetime.strptime(date, format))
TypeError: strptime() argument 1 must be str, not bytes
When we hit a problem like this we can either perform the conversion—in this
case, by passing date.decode("utf8")—or carefully work back to find out where
and why the variable is a bytes object rather than a str, and fix the problem at
its source.
When we pass a string where bytes are expected the error message is somewhat less obvious, and differs between Python 3.0 and 3.1. For example, in
Python 3.0:
Traceback (most recent call last):
File "", line 2139, in <module>
TypeError: expected an object with a buffer interface
In Python 3.1 the error message’s text has been slightly improved:
Traceback (most recent call last):
File "", line 2139, in <module>
TypeError: 'str' does not have the buffer interface
In both cases the problem is that we are passing a string when a bytes, bytearray, or similar object is expected. We can either perform the conversion—in
this case by passing info.encode("utf8")—or work back to find the source of the
problem and fix it there.
Python 3.0 introduced support for exception chaining—this means that an exception that is raised in response to another exception can contain the details of
the original exception. When a chained exception goes uncaught the traceback
includes not just the uncaught exception, but also the exception that caused it
(providing it was chained). The approach to debugging chained exceptions is almost the same as before: We start at the end and work backward until we find
the problem in our own code. However, rather than doing this just for the last
exception, we might then repeat the process for each chained exception above
it, until we get to the problem’s true origin.
We can take advantage of exception chaining in our own code—for example, if
we want to use a custom exception class but still want the underlying problem
to be visible.
class InvalidDataError(Exception): pass
def process(data):
i = int(data)
except ValueError as err:
raise InvalidDataError("Invalid data received") from err
Here, if the int() conversion fails, a ValueError is raised and caught. We
then raise our custom exception, but with from err, which creates a chained
Chapter 9. Debugging, Testing, and Profiling
exception, our own, plus the one in err. If the InvalidDataError exception is
raised and not caught, the resulting traceback will look something like this:
Traceback (most recent call last):
File "", line 249, in process
i = int(data)
ValueError: invalid literal for int() with base 10: '17.5 '
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "", line 288, in <module>
File "", line 283, in process
raise InvalidDataError("Invalid data received") from err
__main__.InvalidDataError: Invalid data received
At the bottom our custom exception and text explain what the problem is, with
the lines above them showing where the exception was raised (line 283), and
where it was caused (line 288). But we can also go back further, into the chained
exception which gives more details about the specific error, and which shows
the line that triggered the exception (249). For a detailed rationale and further
information about chained exceptions, see PEP 3134.
Scientific Debugging
If our program runs but does not have the expected or desired behavior then
we have a bug—a logical error—that we must eliminate. The best way to
eliminate such errors is to prevent them from occurring in the first place by
using TDD (Test Driven Development). However, some bugs will always get
through, so even with TDD, debugging is still a necessary skill to learn.
In this subsection we will outline an approach to debugging based on the scientific method. The approach is explained in sufficient detail that it might appear to be too much work for tackling a “simple” bug. However, by consciously
following the process we will avoid wasting time with “random” debugging, and
after awhile we will internalize the process so that we can do it unconsciously,
and therefore very quickly.★
To be able to kill a bug we must be able to do the following.
1. Reproduce the bug.
2. Locate the bug.
The ideas used in this subsection were inspired by the Debugging chapter in the book Code
Complete by Steve McConnell, ISBN 0735619670.
3. Fix the bug.
4. Test the fix.
Reproducing the bug is sometimes easy—it always occurs on every run; and
sometimes hard—it occurs intermittently. In either case we should try to
reduce the bug’s dependencies, that is, find the smallest input and the least
amount of processing that can still produce the bug.
Once we are able to reproduce the bug, we have the data—the input data and
options, and the incorrect results—that are needed so that we can apply the
scientific method to finding and fixing it. The method has three steps.
1. Think up an explanation—a hypothesis—that reasonably accounts for
the bug.
2. Create an experiment to test the hypothesis.
3. Run the experiment.
Running the experiment should help to locate the bug, and should also give us
insight into its solution. (We will return to how to create and run an experiment
shortly.) Once we have decided how to kill the bug—and have checked our code
into our version control system so that we can revert the fix if necessary—we
can write the fix.
Once the fix is in place we must test it. Naturally, we must test to see if the bug
it is intended to fix has gone away. But this is not sufficient; after all, our fix
may have solved the bug we were concerned about, but the fix might also have
introduced another bug, one that affects some other aspect of the program.
So in addition to testing the bugfix, we must also run all of the program’s
tests to increase our confidence that the bugfix did not have any unwanted
side effects.
Some bugs have a particular structure, so whenever we fix a bug it is always
worth asking ourselves if there are other places in the program or its modules
that might have similar bugs. If there are, we can check to see if we already
have tests that would reveal the bugs if they were present, and if not, we
should add such tests, and if that reveals bugs, then we must tackle them as
described earlier.
Now that we have a good overview of the debugging process, we will focus
in on just how we create and run experiments to test our hypotheses. We
begin with trying to isolate the bug. Depending on the nature of the program
and of the bug, we might be able to write tests that exercise the program, for
example, feeding it data that is known to be processed correctly and gradually
changing the data so that we can find exactly where processing fails. Once
we have an idea of where the problem lies—either due to testing or based on
reasoning—we can test our hypotheses.
Chapter 9. Debugging, Testing, and Profiling
What kind of hypothesis might we think up? Well, it could initially be as simple as the suspicion that a particular function or method is returning erroneous
data when certain input data and options are used. Then, if this hypothesis
proves correct, we can refine it to be more specific—for example, identifying a
particular statement or suite in the function that we think is doing the wrong
computation in certain cases.
To test our hypothesis we need to check the arguments that the function receives and the values of its local variables and the return value, immediately
before it returns. We can then run the program with data that we know produces errors and check the suspect function. If the arguments coming into the
function are not what we expect, then the problem is likely to be further up
the call stack, so we would now begin the process again, this time suspecting
the function that calls the one we have been looking at. But if all the incoming
arguments are always valid, then we must look at the local variables and the
return value. If these are always correct then we need to come up with a new
hypothesis, since the suspect function is behaving correctly. But if the return
value is wrong, then we know that we must investigate the function further.
In practice, how do we conduct an experiment, that is, how do we test the hypothesis that a particular function is misbehaving? One way to start is to
“execute” the function mentally—this is possible for many small functions and
for larger ones with practice, and has the additional benefit that it familiarizes
us with the function’s behavior. At best, this can lead to an improved or more
specific hypothesis—for example, that a particular statement or suite is the
site of the problem. But to conduct an experiment properly we must instrument the program so that we can see what is going on when the suspect function is called.
There are two ways to instrument a program—intrusively, by inserting print()
statements; or (usually) non-intrusively, by using a debugger. Both approaches
are used to achieve the same end and both are valid, but some programmers
have a strong preference for one or the other. We’ll briefly describe both
approaches, starting with the use of print() statements.
When using print() statements, we can start by putting a print() statement
right at the beginning of the function and have it print the function’s arguments. Then, just before the (or each) return statement (or at the end of the
function if there is no return statement), add print(locals(), "\n"). The builtin locals() function returns a dictionary whose keys are the names of the local
variables and whose values are the variables’ values. We can of course simply
print the variables we are specifically interested in instead. Notice that we
added an extra newline—we should also do this in the first print() statement
so that a blank line appears between each set of variables to aid clarity. (An
alternative to inserting print() statements directly is to use some kind of logging decorator such as the one we created in Chapter 8; 358 ➤.)
If when we run the instrumented program we find that the arguments are
correct but that the return value is in error, we know that we have located the
source of the bug and can further investigate the function. If looking carefully
at the function doesn’t suggest where the problem lies, we can simply insert
a new print(locals(), "\n") statement right in the middle. After running the
program again we should now know whether the problem arises in the first
or second half of the function, and can put a print(locals(), "\n") statement
in the middle of the relevant half, repeating the process until we find the
statement where the error is caused. This will very quickly get us to the point
where the problem occurs—and in most cases locating the problem is half of
the work needed to solve it.
The alternative to adding print() statements is to use a debugger. Python
has two standard debuggers. One is supplied as a module (pdb), and can be
used interactively in the console—for example, python3 -m pdb
(On Windows, of course, we would replace python3 with something like
C:\Python31\python.exe.) However, the easiest way to use it is to add import pdb
in the program itself, and add the statement pdb.set_trace() as the first statement of the function we want to examine. When the program is run, pdb stops
it immediately after the pdb.set_trace() call, and allows us to step through the
program, set breakpoints, and examine variables.
Here is an example run of a program that has been instrumented by having
the import pdb statement added to its imports, and by having pdb.set_trace()
added as the first statement inside its calculate_median() function. (What we
have typed is shown in bold, although where we typed Enter is not indicated.)
python3 sum.dat
-> numbers = sorted(numbers)
(Pdb) s
-> middle = len(numbers) // 2
-> median = numbers[middle]
-> if len(numbers) % 2 == 0:
-> return median
(Pdb) p middle, median, numbers
(8, 5.0, [-17.0, -9.5, 0.0, 1.0, 3.0, 4.0, 4.0, 5.0, 5.0, 5.0, 5.5,
6.0, 7.0, 7.0, 8.0, 9.0, 17.0])
(Pdb) c
Chapter 9. Debugging, Testing, and Profiling
Commands are given to pdb by entering their name and pressing Enter at the
(Pdb) prompt. If we just press Enter on its own the last command is repeated.
So here we typed s (which means step, i.e., execute the statement shown), and
then repeated this (simply by pressing Enter), to step through the statements
in the calculate_median() function. Once we reached the return statement we
printed out the values that interested us using the p (print) command. And
finally we continued to the end using the c (continue) command. This tiny
example should give a flavor of pdb, but of course the module has a lot more
functionality than we have shown here.
It is much easier to use pdb on an instrumented program as we have done here
than on an uninstrumented one. But since this requires us to add an import
and a call to pdb.set_trace(), it would seem that using pdb is just as intrusive
as using print() statements, although it does provide useful facilities such
as breakpoints.
The other standard debugger is IDLE, and just like pdb, it supports single
stepping, breakpoints, and the examination of variables. IDLE’s debugger
window is shown in Figure 9.1, and its code editing window with breakpoints
and the current line highlighted is shown in Figure 9.2.
Figure 9.1 IDLE’s debugger window showing the call stack and the current local variables
One great advantage IDLE has over pdb is that there is no need to instrument
our code—IDLE is smart enough to debug our code as it stands, so it isn’t
intrusive at all.
Unfortunately, at the time of this writing, IDLE is rather weak when it comes
to running programs that require command-line arguments. The only way to
do this appears to be to run IDLE from a console with the required arguments,
Figure 9.2 An IDLE code editing window during debugging
for example, idle3 -d -r sum.dat. The -d argument tells IDLE to
start debugging immediately and the -r argument tells it to run the following
program with any arguments that follow it. However, for programs that don’t
require command-line arguments (or where we are willing to edit the code
to put them in manually to make debugging easier), IDLE is quite powerful
and convenient to use. (Incidentally, the code shown in Figure 9.2 does have a
bug—middle + 1 should be middle - 1.)
Debugging Python programs is no harder than debugging in any other
language—and it is easier than for compiled languages since there is no build
step to go through after making changes. And if we are careful to use the scientific method it is usually quite straightforward to locate bugs, although fixing
them is another matter. Ideally, though, we want to avoid as many bugs as possible in the first place. And apart from thinking deeply about our design and
writing our code with care, one of the best ways to prevent bugs is to use TDD,
a topic we will introduce in the next section.
Unit Testing
Writing tests for our programs—if done well—can help reduce the incidence
of bugs and can increase our confidence that our programs behave as expected.
But in general, testing cannot guarantee correctness, since for most nontrivial
programs the range of possible inputs and the range of possible computations
is so vast that only the tiniest fraction of them could ever be realistically
tested. Nonetheless, by carefully choosing what we test we can improve the
quality of our code.
A variety of different kinds of testing can be done, such as usability testing,
functional testing, and integration testing. But here we will concern ourselves
Chapter 9. Debugging, Testing, and Profiling
purely with unit testing—testing individual functions, classes, and methods,
to ensure that they behave according to our expectations.
A key point of TDD, is that when we want to add a feature—for example, a new
method to a class—we first write a test for it. And of course this test will fail
since we haven’t written the method. Now we write the method, and once it
passes the test we can then rerun all the tests to make sure our addition hasn’t
had any unexpected side effects. Once all the tests run (including the one we
added for the new feature), we can check in our code, reasonably confident that
it does what we expect—providing of course that our test was adequate.
For example, if we want to write a function that inserts a string at a particular
index position, we might start out using TDD like this:
def insert_at(string, position, insert):
"""Returns a copy of string with insert inserted at the position
>>> string = "ABCDE"
>>> result = []
>>> for i in range(-2, len(string) + 2):
result.append(insert_at(string, i, "-"))
>>> result[:5]
['ABC-DE', 'ABCD-E', '-ABCDE', 'A-BCDE', 'AB-CDE']
>>> result[5:]
['ABC-DE', 'ABCD-E', 'ABCDE-', 'ABCDE-']
return string
For functions or methods that don’t return anything (they actually return None),
we normally give them a suite consisting of pass, and for those whose return
value is used we either return a constant (say, 0) or one of the arguments,
unchanged—which is what we have done here. (In more complex situations it
may be more useful to return fake objects—third-party modules that provide
“mock” objects are available for such cases.)
When the doctest is run it will fail, listing each of the strings ('ABCD-EF',
'ABCDE-F', etc.) that it expected, and the strings it actually got (all of which
are 'ABCDEF'). Once we are satisfied that the doctest is sufficient and correct,
we can write the body of the function, which in this case is simply return
string[:position] + insert + string[position:]. (And if we wrote return
string[:position] + insert, and then copied and pasted string[:position] at
the end to save ourselves some typing, the doctest will immediately reveal the
Python’s standard library provides two unit testing modules, doctest, which
we have already briefly seen here and earlier (in Chapter 5; 202 ➤, and Chapter 6; 247 ➤), and unittest. In addition, there are third-party testing tools for
Unit Testing
Python. Two of the most notable are nose (,
which aims to be more comprehensive and useful than the standard unittest module, while still being compatible with it, and py.test (codespeak.
net/py/dist/test/test.html)—this takes a somewhat different approach to
unittest, and tries as much as possible to eliminate boilerplate test code. Both
of these third-party tools support test discovery, so there is no need to write an
overarching test program—since they will search for tests themselves. This
makes it easy to test an entire tree of code or just a part of the tree (e.g., just
those modules that have been worked on). For those serious about testing it
is worth investigating both of these third-party modules (and any others that
appeal), before deciding which testing tools to use.
Creating doctests is straightforward: We write the tests in the module, function, class, and methods’ docstrings, and for modules, we simply add three lines
at the end of the module:
if __name__ == "__main__":
import doctest
If we want to use doctests inside programs, that is also possible. For example,
the program whose modules are covered later (in Chapter 14) has
doctests for its functions, but it ends with this code:
if __name__ == "__main__":
This simply calls the program’s main() function, and does not execute the
program’s doctests. To exercise the program’s doctests there are two approaches we can take. One is to import the doctest module and then run the
program—for example, at the console, python3 -m doctest (on Windows, replacing python3 with something like C:\Python31\python.exe). If all
the tests run fine there is no output, so we might prefer to execute python3 -m
doctest -v instead, since this will list every doctest that is executed,
and provide a summary of results at the end.
Another way to execute doctests is to create a separate test program using
the unittest module. The unittest module is conceptually modeled on Java’s
JUnit unit testing library and is used to create test suites that contain test
cases. The unittest module can create test cases based on doctests, without
having to know anything about what the program or module contains, apart
from the fact that it has doctests. So to make a test suite for the
program, we can create the following simple program (which we have called
import doctest
import unittest
import blocks
Chapter 9. Debugging, Testing, and Profiling
suite = unittest.TestSuite()
runner = unittest.TextTestRunner()
Note that there is an implicit restriction on the names of our programs if we
take this approach: They must have names that are valid module names, so
a program called cannot have a test like this written for
it because import convert-incidents is not valid since hyphens are not legal in
Python identifiers. (It is possible to get around this, but the easiest solution
is to use program filenames that are also valid module names, for example,
replacing hyphens with underscores.)
The structure shown here—create a test suite, add one or more test cases or
test suites, run the overarching test suite, and output the results—is typical
of unittest-based tests. When run, this particular example produces the
following output:
---------------------------------------------------------------------Ran 3 tests in 0.244s
<unittest._TextTestResult run=3 errors=0 failures=0>
Each time a test case is executed a period is output (hence the three periods
at the beginning of the output), then a line of hyphens, and then the test
summary. (Naturally, there is a lot more output if any tests fail.)
If we are making the effort to have separate tests (typically one for each program and module we want to test), then rather than using doctests we might
prefer to directly use the unittest module’s features—especially if we are used
to the JUnit approach to testing. The unittest module keeps our tests separate
from our code—this is particularly useful for larger projects where test writers
and developers are not necessarily the same people. Also, unittest unit tests
are written as stand-alone Python modules, so they are not limited by what we
can comfortably and sensibly write inside a docstring.
The unittest module defines four key concepts. A test fixture is the term used
to describe the code necessary to set up a test (and to tear it down, that is,
clean up, afterward). Typical examples are creating an input file for the test to
use and at the end deleting the input file and the resultant output file. A test
suite is a collection of test cases and a test case is the basic unit of testing—test
suites are collections of test cases or of other test suites—we’ll see practical
examples of these shortly. A test runner is an object that executes one or more
test suites.
Unit Testing
Typically, a test suite is made by creating a subclass of unittest.TestCase,
where each method that has a name beginning with “test” is a test case. If
we need any setup to be done, we can do it in a method called setUp(); similarly, for any cleanup we can implement a method called tearDown(). Within the
tests there are a number of unittest.TestCase methods that we can make use
of, including assertTrue(), assertEqual(), assertAlmostEqual() (useful for testing floating-point numbers), assertRaises(), and many more, including many
inverses such as assertFalse(), assertNotEqual(), failIfEqual(), failUnlessEqual(), and so on. ex-
411 ➤
The unittest module is well documented and has a lot of functionality, but here
we will just give a flavor of its use by reviewing a very simple test suite. The
example we will use is the solution to one of the exercises given at the end of
Chapter 8. The exercise was to create an Atomic module which could be used
as a context manager to ensure that either all of a set of changes is applied to
a list, set, or dictionary—or none of them are. The module provided
as an example solution uses 30 lines of code to implement the Atomic class,
and has about 100 lines of module doctests. We will create the
module to replace the doctests with unittest tests so that we can then delete
the doctests and leave free of any code except that needed to provide
its functionality.
Before diving into writing the test module, we need to think about what tests
are needed. We will need to test three different kinds of data type: lists, sets,
and dictionaries. For lists we need to test appending and inserting an item,
deleting an item, and changing an item’s value. For sets we must test adding
and discarding an item. And for dictionaries we must test inserting an item,
changing an item’s value, and deleting an item. Also, we must test that in the
case of failure, none of the changes are applied.
Structurally, testing the different data types is essentially the same, so we will
only write the test cases for testing lists and leave the others as an exercise.
The module must import both the unittest module and the
Atomic module that it is designed to test.
When creating unittest files, we usually create modules rather than programs,
and inside each module we define one or more unittest.TestCase subclasses.
In the case of the module, it defines a single unittest.TestCase
subclass, TestAtomic (which we will review shortly), and ends with the following
two lines:
if __name__ == "__main__":
Thanks to these lines, the module can be run stand-alone. And of course, it
could also be imported and run from another test program—something that
makes sense if this is just one test suite among many.
Chapter 9. Debugging, Testing, and Profiling
If we want to run the module from another test program we can
write a program that is similar to the one we used to execute doctests using the
unittest module. For example:
import unittest
import test_Atomic
suite = unittest.TestLoader().loadTestsFromTestCase(
runner = unittest.TextTestRunner()
Here, we have created a single suite by telling the unittest module to read the
test_Atomic module and to use each of its test*() methods (test_list_success()
and test_list_fail() in this example, as we will see in a moment), as test
We will now review the implementation of the TestAtomic class. Unusually for
subclasses generally, although not for unittest.TestCase subclasses, there is no
need to implement the initializer. In this case we will need a setup method, but
not a teardown method. And we will implement two test cases.
def setUp(self):
self.original_list = list(range(10))
We have used the unittest.TestCase.setUp() method to create a single piece of
test data.
def test_list_succeed(self):
items = self.original_list[:]
with Atomic.Atomic(items) as atomic:
atomic.insert(2, -915)
del atomic[5]
atomic[4] = -782
atomic.insert(0, -9)
[-9, 0, 1, -915, 2, -782, 5, 6, 7, 8, 9, 1999])
This test case is used to test that all of a set of changes to a list are correctly
applied. The test performs an append, an insertion in the middle, an insertion
at the beginning, a deletion, and a change of a value. While by no means
comprehensive, the test does at least cover the basics.
The test should not raise an exception, but if it does the unittest.TestCase
base class will handle it by turning it into an appropriate error message. At
the end we expect the items list to equal the literal list included in the test
rather than the original list. The unittest.TestCase.assertEqual() method can
Unit Testing
compare any two Python objects, but its generality means that it cannot give
particularly informative error messages.
From Python 3.1, the unittest.TestCase class has many more methods, including many data-type-specific assertion methods. Here is how we could write the
assertion using Python 3.1:
[-9, 0, 1, -915, 2, -782, 5, 6, 7, 8, 9, 1999])
If the lists are not equal, since the data types are known, the unittest module
is able to give more precise error information, including where the lists differ.
def test_list_fail(self):
def process():
nonlocal items
with Atomic.Atomic(items) as atomic:
atomic.insert(2, -915)
del atomic[5]
atomic[4] = -782
atomic.poop() # Typo
items = self.original_list[:]
self.assertRaises(AttributeError, process)
self.assertEqual(items, self.original_list)
To test the failure case, that is, where an exception is raised while doing atomic
processing, we must test that the list has not been changed and also that an
appropriate exception has been raised. To check for an exception we use the
unittest.TestCase.assertRaises() method, and in the case of Python 3.0, we
pass it the exception we expect to get and a callable object that should raise the
exception. This forces us to encapsulate the code we want to test, which is why
we had to create the process() inner function shown here.
In Python 3.1 the unittest.TestCase.assertRaises() method can be used as a
context manager, so we are able to write our test in a much more natural way:
def test_list_fail(self):
items = self.original_list[:]
with self.assertRaises(AttributeError):
with Atomic.Atomic(items) as atomic:
atomic.insert(2, -915)
del atomic[5]
atomic[4] = -782
atomic.poop() # Typo
self.assertListEqual(items, self.original_list)
Chapter 9. Debugging, Testing, and Profiling
Here we have written the test code directly in the test method without the
need for an inner function, instead using unittest.TestCase.assertRaised() as a
context manager that expects the code to raise an AttributeError. We have also
used Python 3.1’s unittest.TestCase.assertListEqual() method at the end.
As we have seen, Python’s test modules are easy to use and are extremely useful, especially if we use TDD. They also have a lot more functionality and features than have been shown here—for example, the ability to skip tests which
is useful to account for platform differences—and they are also well documented. One feature that is missing—and which nose and py.test provide—is test
discovery, although this feature is expected to appear in a later Python version
(perhaps as early as Python 3.2).
If a program runs very slowly or consumes far more memory than we expect,
the problem is most often due to our choice of algorithms or data structures, or
due to our doing an inefficient implementation. Whatever the reason for the
problem, it is best to find out precisely where the problem lies rather than just
inspecting our code and trying to optimize it. Randomly optimizing can cause
us to introduce bugs or to speed up parts of our program that actually have no
effect on the program’s overall performance because the improvements are not
in places where the interpreter spends most of its time.
Before going further into profiling, it is worth noting a few Python programming habits that are easy to learn and apply, and that are good for performance. None of the techniques is Python-version-specific, and all of them
are perfectly sound Python programming style. First, prefer tuples to lists
when a read–only sequence is needed. Second, use generators rather than
creating large tuples or lists to iterate over. Third, use Python’s built-in data
structures—dicts, lists, and tuples—rather than custom data structures
implemented in Python, since the built-in ones are all very highly optimized.
Fourth, when creating large strings out of lots of small strings, instead of concatenating the small strings, accumulate them all in a list, and join the list of
strings into a single string at the end. Fifth and finally, if an object (including a
function or method) is accessed a large number of times using attribute access
(e.g., when accessing a function in a module), or from a data structure, it may
be better to create and use a local variable that refers to the object to provide
faster access.
Python’s standard library provides two modules that are particularly useful
when we want to investigate the performance of our code. One of these is the
timeit module—this is useful for timing small pieces of Python code, and can be
used, for example, to compare the performance of two or more implementations
of a particular function or method. The other is the cProfile module which can
be used to profile a program’s performance—it provides a detailed breakdown
of call counts and times and so can be used to find performance bottlenecks.★
To give a flavor of the timeit module, we will look at a small example. Suppose
we have three functions, function_a(), function_b(), and function_c(), all of
which perform the same computation, but each using a different algorithm.
If we put all these functions into a module (or import them), we can run them
using the timeit module to see how they compare. Here is the code that we
would use at the end of the module:
if __name__ == "__main__":
repeats = 1000
for function in ("function_a", "function_b", "function_c"):
t = timeit.Timer("{0}(X, Y)".format(function),
"from __main__ import {0}, X, Y".format(function))
sec = t.timeit(repeats) / repeats
print("{function}() {sec:.6f} sec".format(**locals()))
The first argument given to the timeit.Timer() constructor is the code we want
to execute and time, in the form of a string. Here, the first time around the loop,
the string is "function_a(X, Y)". The second argument is optional; again it is a
string to be executed, this time before the code to be timed so as to provide some
setup. Here we have imported from the __main__ (i.e., this) module the function
we want to test, plus two variables that are passed as input data (X and Y), and
that are available as global variables in the module. We could just as easily
have imported the function and data from a different module.
When the timeit.Timer object’s timeit() method is called, it will first execute
the constructor’s second argument—if there was one—to set things up, and
then it will execute the constructor’s first argument—and time how long the
execution takes. The timeit.Timer.timeit() method’s return value is the time
taken in seconds, as a float. By default, the timeit() method repeats 1 million
times and returns the total seconds for all these executions, but in this particular case we needed only 1 000 repeats to give us useful results, so we specified
the repeat count explicitly. After timing each function we divide the total by
the number of repeats to get its mean (average) execution time and print the
function’s name and execution time on the console.
function_a() 0.001618 sec
function_b() 0.012786 sec
function_c() 0.003248 sec
In this example, function_a() is clearly the fastest—at least with the input
data that was used. In some situations—for example, where performance can
The cProfile module is usually available for CPython interpreters, but is not always available for
others. All Python libraries should have the pure Python profile module which provides the same
API as the cProfile module, and does the same job, only more slowly.
Chapter 9. Debugging, Testing, and Profiling
vary considerably depending on the input data—we might have to test each
function with multiple sets of input data to cover a representative set of cases
and then compare the total or average execution times.
It isn’t always convenient to instrument our code to get timings, and so the
timeit module provides a way of timing code from the command line. For
example, to time function_a() from the module, we would enter
the following in the console: python3 -m timeit -n 1000 -s "from MyModule import
function_a, X, Y" "function_a(X, Y)". (As usual, for Windows, we must replace
python3 with something like C:\Python31\python.exe.) The -m option is for the
Python interpreter and tells it to load the specified module (in this case timeit)
and the other options are handled by the timeit module. The -n option specifies
the repetition count, the -s option specifies the setup, and the last argument
is the code to execute and time. After the command has finished it prints its
results on the console, for example:
1000 loops, best of 3: 1.41 msec per loop
We can easily then repeat the timing for the other two functions so that we can
compare them all.
The cProfile module (or the profile module—we will refer to them both as the
cProfile module) can also be used to compare the performance of functions
and methods. And unlike the timeit module that just provides raw timings,
the cProfile module shows precisely what is being called and how long each
call takes. Here’s the code we would use to compare the same three functions
as before:
if __name__ == "__main__":
for function in ("function_a", "function_b", "function_c"):"for i in range(1000): {0}(X, Y)"
We must put the number of repeats inside the code we pass to the function, but we don’t need to do any setup since the module function uses introspection to find the functions and variables we want to use.
There is no explicit print() statement since by default the function prints its output on the console. Here are the results for all the functions
(with some irrelevant lines omitted and slightly reformatted to fit the page):
1003 function calls in 1.661 CPU seconds
ncalls tottime percall cumtime percall filename:lineno(function)
1.661 <string>:1(<module>)
1.661 {built-in method exec}
5132003 function calls in 22.700 CPU seconds
ncalls tottime percall cumtime percall filename:lineno(function)
0.487 22.700 22.700 <string>:1(<module>)
0.000 22.213
0.000 {built-in method bisect_left}
0.000 22.700 22.700 {built-in method exec}
0.000 {built-in method len}
1000 15.149
0.015 22.196
0.022 {built-in method sorted}
5129003 function calls in 12.987 CPU seconds
ncalls tottime percall cumtime percall filename:lineno(function)
0.205 12.987 12.987 <string>:1(<module>)
0.006 12.782
0.000 12.987 12.987 {built-in method exec}
The ncalls (“number of calls”) column lists the number of calls to the specified
function (listed in the filename:lineno(function) column). Recall that we repeated the calls 1 000 times, so we must keep this in mind. The tottime (“total
time”) column lists the total time spent in the function, but excluding time
spent inside functions called by the function. The first percall column lists
the average time of each call to the function (tottime // ncalls). The cumtime
(“cumulative time”) column lists the time spent in the function and includes the
time spent inside functions called by the function. The second percall column
lists the average time of each call to the function, including functions called
by it.
This output is far more enlightening than the timeit module’s raw timings. We
can immediately see that both function_b() and function_c() use generators
that are called more than 5 000 times, making them both at least ten times
slower than function_a(). Furthermore, function_b() calls more functions generally, including a call to the built-in sorted() function, and this makes it almost
twice as slow as function_c(). Of course, the timeit() module gave us sufficient
information to see these differences in timing, but the cProfile module allows
us to see the details of why the differences are there in the first place.
Just as the timeit module allows us to time code without instrumenting it, so
does the cProfile module. However, when using the cProfile module from the
command line we cannot specify exactly what we want executed—it simply
executes the given program or module and reports the timings of everything.
The command line to use is python3 -m cProfile, and the
output produced is in the same format as we saw earlier; here is an extract
slightly reformatted and with most lines omitted:
Chapter 9. Debugging, Testing, and Profiling
10272458 function calls (10272457 primitive calls) in 37.718 CPU secs
ncalls tottime percall cumtime percall filename:lineno(function)
37.718 <string>:1(<module>)
37.717 <string>:12(<module>)
0.002 <string>:20(function_a)
0.023 <string>:27(function_b)
0.000 <string>:28(<genexpr>)
0.013 <string>:35(function_c)
0.000 <string>:36(<genexpr>)
In cProfile terminology, a primitive call is a nonrecursive function call.
Using the cProfile module in this way can be useful for identifying areas that
are worth investigating further. Here, for example, we can clearly see that
function_b() takes a long time. But how do we drill down into the details? We
could instrument the program by replacing calls to function_b() with cProfile.
run("function_b()"). Or we could save the complete profile data and analyze it
using the pstats module. To save the profile we must modify our command line
slightly: python3 -m cProfile -o profileDataFile We can then
analyze the profile data, for example, by starting IDLE, importing the pstats
module, and giving it the saved profileDataFile, or by using pstats interactively at the console. Here’s a very short example console session that has been
tidied up slightly to fit on the page, and with our input shown in bold:
$ python3 -m cProfile -o profile.dat
$ python3 -m pstats
Welcome to the profile statistics browser.
% read profile.dat
profile.dat% callers function_b
Random listing order was used
List reduced from 44 to 1 due to restriction <'function_b'>
Function was called by...
ncalls tottime cumtime
<string>:27(function_b) <- 1000
0.011 22.251 <string>:12(<module>)
profile.dat% callees function_b
Random listing order was used
List reduced from 44 to 1 due to restriction <'function_b'>
Function called...
ncalls tottime cumtime
<string>:27(function_b) ->
0.005 built-in method bisect_left
0.001 built-in method len
1000 15.297 22.234 built-in method sorted
profile.dat% quit
Type help to get the list of commands, and help followed by a command name
for more information on the command. For example, help stats will list what
arguments can be given to the stats command. Other tools are available
that can provide a graphical visualization of the profile data, for example,
RunSnakeRun (, which depends on
the wxPython GUI library.
Using the timeit and cProfile modules we can identify areas of our code that
might be taking more time than expected, and using the cProfile module, we
can find out exactly where the time is being taken.
In general, Python’s reporting of syntax errors is very accurate, with the line
and position in the line being correctly identified. The only cases where this
doesn’t work well are when we forget a closing parenthesis, bracket, or brace,
in which case the error is normally reported as being on the next nonblank line.
Fortunately, syntax errors are almost always easy to see and to fix.
If an unhandled exception is raised, Python will terminate and output a traceback. Such tracebacks can be intimidating for end-users, but provide useful
information to us as programmers. Ideally, we should always handle every
type of exception that we believe our program can raise, and where necessary
present the problem to the user in the form of an error message, message box,
or log message—but not as a raw traceback. However, we should avoid using
the catchall except: exception handler—if we want to handle all exceptions
(e.g., at the top level), then we can use except Exception as err, and always report
err, since silently handling exceptions can lead to programs failing in subtle
and unnoticed ways (such as corrupting data) later on. And during development, it is probably best not to have a top-level exception handler at all and to
simply have the program crash with a traceback.
Debugging need not be—and should not be—a hit and miss affair. By narrowing down the input necessary to reproduce the bug to the bare minimum, by
carefully hypothesizing what the problem is, and then testing the hypothesis
by experiment—using print() statements or a debugger—we can often locate
the source of the bug quite quickly. And if our hypothesis has successfully led
us to the bug, it is likely to also be helpful in devising a solution.
For testing, both the doctest and the unittest modules have their own particular virtues. Doctests tend to be particularly convenient and useful for small
libraries and modules since well-chosen tests can easily both illustrate and
exercise boundary as well as common cases, and of course, writing doctests is
convenient and easy. On the other hand, since unit tests are not constrained to
be written inside docstrings and are written as separate stand-alone modules,
they are usually a better choice when it comes to writing more complex and
Chapter 9. Debugging, Testing, and Profiling
sophisticated tests, especially tests that require setup and teardown (cleanup).
For larger projects, using the unittest module (or a third-party unit testing
module) keeps the tests and tested programs and modules separate and is generally more flexible and powerful than using doctests.
If we hit performance problems, the cause is most often our own code, and in
particular our choice of algorithms and data structures, or some inefficiency in
our implementation. When faced with such problems, it is always wise to find
out exactly where the performance bottleneck is, rather than to guess and end
up spending time optimizing something that doesn’t actually improve performance. Python’s timeit module can be used to get raw timings of functions or
arbitrary code snippets, and so is particularly useful for comparing alternative
function implementations. And for in-depth analysis, the cProfile module provides both timing and call count information so that we can identify not only
which functions take the most time, but also what functions they in turn call.
Overall, Python has excellent support for debugging, testing, and profiling,
right out of the box. However, especially for large projects, it is worth considering some of the third-party testing tools, since they may offer more functionality and convenience than the standard library’s testing modules provide.
● Using the Multiprocessing Module
● Using the Threading Module
Processes and Threading
With the advent of multicore processors as the norm rather than the exception,
it is more tempting and more practical than ever before to want to spread
the processing load so as to get the most out of all the available cores. There
are two main approaches to spreading the workload. One is to use multiple
processes and the other is to use multiple threads. This chapter shows how to
use both approaches.
Using multiple processes, that is, running separate programs, has the advantage that each process runs independently. This leaves all the burden of handling concurrency to the underlying operating system. The disadvantage is
that communication and data sharing between the invoking program and the
separate processes it invokes can be inconvenient. On Unix systems this can
be solved by using the exec and fork paradigm, but for cross-platform programs other solutions must be used. The simplest, and the one shown here, is
for the invoking program to feed data to the processes it runs and leave them
to produce their output independently. A more flexible approach that greatly
simplifies two-way communication is to use networking. Of course, in many
situations such communication isn’t needed—we just need to run one or more
other programs from one orchestrating program.
An alternative to handing off work to independent processes is to create a
threaded program that distributes work to independent threads of execution.
This has the advantage that we can communicate simply by sharing data (providing we ensure that shared data is accessed only by one thread at a time), but
leaves the burden of managing concurrency squarely with the programmer.
Python provides good support for creating threaded programs, minimizing the
work that we must do. Nonetheless, multithreaded programs are inherently
more complex than single-threaded programs and require much more care in
their creation and maintenance.
In this chapter’s first section we will create two small programs. The first program is invoked by the user and the second program is invoked by the first pro439
Chapter 10. Processes and Threading
gram, with the second program invoked once for each separate process that is
required. In the second section we will begin by giving a bare-bones introduction to threaded programming. Then we will create a threaded program that
has the same functionality as the two programs from the first section combined
so as to provide a contrast between the multiple processes and the multiple
threads approaches. And then we will review another threaded program, more
sophisticated than the first, that both hands off work and gathers together all
the results.
Using the Multiprocessing Module
In some situations we already have programs that have the functionality we
need but we want to automate their use. We can do this by using Python’s subprocess module which provides facilities for running other programs, passing
any command-line options we want, and if desired, communicating with them
using pipes. We saw one very simple example of this in Chapter 5 when we
used the function to clear the console in a platform-specific
way. But we can also use these facilities to create pairs of “parent–child” programs, where the parent program is run by the user and this in turn runs as
many instances of the child program as necessary, each with different work to
do. It is this approach that we will cover in this section.
In Chapter 3 we showed a very simple program,, that searches
for a word specified on the command line in the files listed after the word. In
this section we will develop a more sophisticated version that can recurse into
subdirectories to find files to read and that can delegate the work to as many
separate child processes as we like. The output is just a list of filenames (with
paths) for those files that contain the specified search word.
The parent program is and the child program is The relationship between the two programs when they are being run
is shown schematically in Figure 10.1.
The heart of is encapsulated by its main() function, which we will
look at in three parts:
def main():
child = os.path.join(os.path.dirname(__file__),
opts, word, args = parse_options()
filelist = get_files(args, opts.recurse)
files_per_process = len(filelist) // opts.count
start, end = 0, files_per_process + (len(filelist) % opts.count)
number = 1
Chapter 10. Processes and Threading
the Python interpreter, but we prefer this approach because it ensures that the
child program uses the same Python interpreter as the parent program.
Once we have the command ready we create a subprocess.Popen object, specifying the command to execute (as a list of strings), and in this case requesting
to write to the process’s standard input. (It is also possible to read a process’s
standard output by setting a similar keyword argument.) We then write the
search word followed by a newline and then every file in the relevant slice of
the file list. The subprocess module reads and writes bytes, not strings, but the
processes it creates always assume that the bytes received from sys.stdin are
strings in the local encoding—even if the bytes we have sent use a different encoding, such as UTF-8 which we have used here. We will see how to get around
this annoying problem shortly. Once the word and the list of files have been
written to the child process, we close its standard input and move on.
It is not strictly necessary to keep a reference to each process (the pipe variable
gets rebound to a new subprocess.Popen object each time through the loop),
since each process runs independently, but we add each one to a list so that we
can make them interruptible. Also, we don’t gather the results together, but
instead we let each process write its results to the console in its own time. This
means that the output from different processes could be interleaved. (You will
get the chance to avoid interleaving in the exercises.)
while pipes:
pipe = pipes.pop()
Once all the processes have started we wait for each child process to finish. This
is not essential, but on Unix-like systems it ensures that we are returned to the
console prompt when all the processes are done (otherwise, we must press Enter
when they are all finished). Another benefit of waiting is that if we interrupt
the program (e.g., by pressing Ctrl+C), all the processes that are still running
will be interrupted and will terminate with an uncaught KeyboardInterrupt
exception—if we did not wait the main program would finish (and therefore not
be interruptible), and the child processes would continue (unless killed by a kill
program or a task manager).
Apart from the comments and imports, here is the complete program. We will look at the program in two parts—with two versions of the first part, the first for any Python 3.x version and the second for
Python 3.1 or later versions:
number = "{0}: ".format(sys.argv[1]) if len(sys.argv) == 2 else ""
stdin =
lines = stdin.decode("utf8", "ignore").splitlines()
word = lines[0].rstrip()
Using the Multiprocessing Module
The program begins by setting the number string to the given number or to
an empty string if we are not debugging. Since the program is running as a
child process and the subprocess module only reads and writes binary data
and always uses the local encoding, we must read sys.stdin’s underlying buffer
of binary data and perform the decoding ourselves.★ Once we have read the
binary data, we decode it into a Unicode string and split it into lines. The child
process then reads the first line, since this contains the search word.
Here are the lines that are different for Python 3.1:
sys.stdin = sys.stdin.detach()
stdin =
lines = stdin.decode("utf8", "ignore").splitlines()
Python 3.1 provides the sys.stdin.detach() method that returns a binary file
object. We then read in all the data, decode it into Unicode using the encoding
of our choice, and then split the Unicode string into lines.
for filename in lines[1:]:
filename = filename.rstrip()
previous = ""
with open(filename, "rb") as fh:
while True:
current =
if not current:
current = current.decode("utf8", "ignore")
if (word in current or
word in previous[-len(word):] +
print("{0}{1}".format(number, filename))
if len(current) != BLOCK_SIZE:
previous = current
except EnvironmentError as err:
print("{0}{1}".format(number, err))
All the lines after the first are filenames (with paths). For each one we open
the relevant file, read it, and print its name if it contains the search word. It is
possible that some of the files might be very large and this could be a problem,
especially if there are 20 child processes running concurrently, all reading big
It is possible that a future version of Python will have a version of the subprocess module that
allows encoding and errors arguments so that we can use our preferred encoding without having
to access sys.stdin in binary mode and do the decoding ourselves. See
Chapter 10. Processes and Threading
files. We handle this by reading each file in blocks, keeping the previous block
read to ensure that we don’t miss cases when the only occurrence of the search
word happens to fall across two blocks. Another benefit of reading in blocks
is that if the search word appears early in the file we can finish with the file
without having read everything, since all we care about is whether the word is
in the file, not where it appears within the file.
91 ➤
The files are read in binary mode, so we must convert each block to a string before we can search it, since the search word is a string. We have assumed that
all the files use the UTF-8 encoding, but this is most likely wrong in some cases.
A more sophisticated program would try to determine the actual encoding and
then close and reopen the file using the correct encoding. As we noted in Chapter 2, at least two Python packages for automatically detecting a file’s encoding
are available from the Python Package Index, (It might
be tempting to decode the search word into a bytes object and compare bytes
with bytes, but that approach is not reliable since some characters have more
than one valid UTF-8 representation.)
The subprocess module offers a lot more functionality than we have needed to
use here, including the ability to provide equivalents to shell backquotes and
shell pipelines, and to the os.system() and spawn functions.
In the next section we will see a threaded version of the program
so that we can compare it with the parent–child processes one. We will also
look at a more sophisticated threaded program that delegates work and then
gathers the results together to have more control over how they are output.
Using the Threading Module
Setting up two or more separate threads of execution in Python is quite
straightforward. The complexity arises when we want separate threads to
share data. Imagine that we have two threads sharing a list. One thread
might start iterating over the list using for x in L and then somewhere in the
middle another