SQLBase Advanced Topics Guide
Advanced Topics Guide
Centura, Centura Ranger, the Centura logo, Centura Web Developer, Gupta, the Gupta
logo, Gupta Powered, the Gupta Powered logo, Fast Facts, Object Nationalizer,
Quest, Quest/Web, QuickObjects, SQL/API, SQLBase, SQLConsole, SQLGateway,
SQLHost, SQLNetwork, SQLRouter, SQLTalk, and Team Object Manager are
trademarks of Centura Software Corporation and may be registered in the United
States of America and/or other countries. SQLWindows is a registered trademark and
TeamWindows, ReportWindows and EditWindows are trademarks exclusively used
and licensed by Centura Software Corporation.
Microsoft, Win32, Windows, Windows NT and Visual Basic are either registered
trademarks or trademarks of Microsoft Corporation in the United States of America
and/or other countries.
IBM and IBM PC are registered trademarks of International Business Machines
Corporation. DB2/ OS/2, and Token Ring are trademarks of International Business
NetWare and Novell are registered trademarks of Novell, Inc. Netware Requester is a
trademark of Novell, Inc.
Java is a trademark of Sun Microsystems Inc.
SQL Designs and the SQL Designs logo are trademarks of SQL Designs, Inc.
All other product or service names mentioned herein are trademarks or registered
trademarks of their respective owners.
Copyright  1997 by Centura Software Corporation. All rights reserved.
Advanced Topics Guide
November 1997
Advanced Topics Guide
In the early days of relational databases there was a certain purist mentality where
relational databases were compared to hierarchical and network databases, and were
viewed as impure because users had detailed knowledge of the data structures used to
store data. Often, users would alter the design and implementation of their programs
to take into account the structure of data storage. Relational database vendors took the
high ground in developing the concept that it was the RDBMS’s responsibility to
efficiently manage the storage and retrieval of data, and it was the responsibility of
the application designer to provide the logical design and schema of the data.
Coming from this purist background I have often resisted the notion of allowing our
users to know about the internals of the database. However, time, experience, and the
realities of relational technology have taught me that a perfect world where data
storage and retrieval is optimally handled by the RDBMS does not exist for us or any
other RDBMS vendor. Our customer base has been very clear in communicating their
need to know more and more about the internals of SQLBase. Our close
communication with customers and consultants have helped us to better understand
the need and the value of clearly explaining and documenting how SQLBase does
what it does.
With version 6 of SQLBase, we have taken the opposite position on exposing the
internals of SQLBase. With SQLConsole 2.0, our customers can learn almost as
much about the internal workings of SQLBase as the engineers that developed it. We
have come to the realization that more knowledge about the internals benefits both
our customers and ourselves. Our customers are able to build more robust, better
performing, and more stable applications with increased knowledge of SQLBase
internals. We benefit from the success of our customers in deploying these robust
applications. We also benefit from receiving feedback from customers whose
knowledge has been enhanced by information we have provided them about SQLBase
internals. From all of this it became clear that both our customers and ourselves win
by producing this book.
Advanced Topics Guide
In 1992 I was approached by Jon Boring of SQL Designs with the idea of producing a
book that would thoroughly document SQLBase internals. This notion was very
exciting because of the great need for such information, and also because of the value
of having this document produced by a group who is very experienced in the logical
and physical design of relational databases. SQL Designs is one of our oldest and
most technically respected business partners. The people at SQL Designs have largeshop, mainframe-class database know-how and have been very effective in
transferring this knowledge to the PC client/server world. They have a deep
understanding of the kinds of information a professional database administrator needs
for database design and administration. SQL Designs have participated in some of the
largest and most successful SQLBase projects and are the preferred SQLBase
consultants for some of our largest SQLBase customers.
In 1991 and 1992, they were involved in several large SQLBase projects. From this
project SQL Designs also became very much aware of the great need for such internal
documentation. I saw the great synergism that could happen if the SQLBase
development staff and SQL Designs pooled their expertise and talents to produce this
From October 1992 to December 1993 SQL Designs embarked upon an intense
research project to learn and understand SQLBase internals. They worked closely
with SQLBase development engineers to understand every aspect of the way
SQLBase operates. The information and ideas were refined and reviewed over and
over again by SQLBase engineering and publications staff for accuracy and clarity.
For a period of time it was as though the people from SQL Designs became a part of
the SQLBase 6.0 development project. In fact, the interaction with SQL Designs
resulted in the decision to implement a new feature in SQLBase that allows the
internal statistics SQLBase uses for its query optimization to be altered for the
purpose of query modeling. This feature and several other features implemented in
SQLBase 6.0 were a direct result of our interaction with SQL Designs.
Through the SQLBase 6.0 beta program we received excellent feedback about the
value of this book. It was the long awaited answer to many of our customers. We
believe this book will be the answer to those who have been thirsting after the
detailed internal knowledge of SQLBase. We hope and believe that this book will
become a vital part of the application development process involved in developing
applications and systems using SQLBase.
Bruce D. Scott
Senior Vice President Research & Development,
DataBase and Connectivity Products
Gupta Corporation
Advanced Topics Guide
About the Authors
About the Authors
The SQLBase Advanced Topics Guide was written by Jon Boring and Dave Gerber of
SQL Designs, Inc. Jon Boring, president and cofounder of SQL Designs, is a
nationally recognized expert in the area of client/server software development
methods and a frequent speaker at industry conferences and trade shows.
Prior to joining SQL Designs, Dave Gerber directed the database administration and
systems programming functions at several Fortune 500 corporations. Dave is a senior
manager at SQL Designs and oversees the database technology group where he
specializes in database administration, physical database design and query optimizers
for SQL databases.
SQL Designs is a database consulting organization which specializes in the client/
server architecture and has extensive experience using the Gupta products. They offer
consulting, applications development, and training in addition to a full range of
support services. Interested parties may contact SQL Designs at:
SQL Designs, Inc.
206 East Pine Street
Lakeland, FL 339801
Telephone: 813.683.3886
Fax: 813.683.3997
Advanced Topics Guide
Advanced Topics Guide
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
1 Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1
ANSI/SPARC three schema architecture . . . . . . . . . . 1-2
Structure of this section. . . . . . . . . . . . . . . . . . . . . . . . 1-4
2 Conceptual Schema Creation
. . . . . . . . . 2-1
Data administration . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
Standards creation and enforcement . . . . . . . . . . 2-2
Data dictionary . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4
Enterprise-wide data planning . . . . . . . . . . . . . . . 2-4
Subject area databases . . . . . . . . . . . . . . . . . . . . 2-5
Logical data modeling . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
Entities and relationships . . . . . . . . . . . . . . . . . . . 2-6
Attributes and domains . . . . . . . . . . . . . . . . . . . . 2-11
Primary and foreign keys . . . . . . . . . . . . . . . . . . 2-11
Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
Normal forms . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-14
Denormalization . . . . . . . . . . . . . . . . . . . . . . . . . 2-18
3 Internal Schema Creation . . . . . . . . . . . . . . 3-1
Building the initial internal schema . . . . . . . . . . . . . . . 3-2
Creating tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
Adding columns to tables . . . . . . . . . . . . . . . . . . . 3-3
Data typing for columns . . . . . . . . . . . . . . . . . . . . 3-3
Null usage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-5
Advanced Topics Guide
Resolving many-to-many relationships . . . . . . . . . . . . 3-6
Adding referential integrity . . . . . . . . . . . . . . . . . . . . . 3-8
Defining keys . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
4 External Schema Creation . . . . . . . . . . . . . 4-1
Using views as the external schema. . . . . . . . . . . . . . 4-2
Creating the initial external schema . . . . . . . . . . . . . . 4-4
Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-5
5 Transaction Definition . . . . . . . . . . . . . . . . . . 5-1
Benefits of defining transactions . . . . . . . . . . . . . . . . . 5-2
Physical database design . . . . . . . . . . . . . . . . . . . 5-2
Data integrity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-2
Defining transactions. . . . . . . . . . . . . . . . . . . . . . . . . . 5-3
Transaction name, number, and description . . . . 5-3
Transaction type and complexity . . . . . . . . . . . . . 5-4
Transaction volumes. . . . . . . . . . . . . . . . . . . . . . . 5-4
Transaction performance requirements . . . . . . . . 5-5
Relative priority . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
Transaction SQL statements . . . . . . . . . . . . . . . . 5-6
6 Refining the Physical Design
. . . . . . . . . . 6-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
Splitting tables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
Vertically splitting large rows . . . . . . . . . . . . . . . . 6-3
Horizontally splitting tables . . . . . . . . . . . . . . . . . . 6-4
Tailoring performance requirements. . . . . . . . . . . 6-5
Referential integrity considerations. . . . . . . . . . . . 6-5
Clustering of table rows. . . . . . . . . . . . . . . . . . . . . . . . 6-7
Direct random access of rows . . . . . . . . . . . . . . . 6-7
Clustering rows together. . . . . . . . . . . . . . . . . . . . 6-8
Indexing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
Column cardinality and selectivity factor. . . . . . . 6-11
Advanced Topics Guide
Composite indexes . . . . . . . . . . . . . . . . . . . . . . . 6-12
Good index candidates . . . . . . . . . . . . . . . . . . . . 6-14
Poor index candidates . . . . . . . . . . . . . . . . . . . . 6-15
Database partitioning . . . . . . . . . . . . . . . . . . . . . . . . 6-16
Free space calculations . . . . . . . . . . . . . . . . . . . . . . 6-18
7 Physical Design Control. . . . . . . . . . . . . . . . 7-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-2
The process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-2
Step one . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
Step two . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-6
8 Database Pages . . . . . . . . . . . . . . . . . . . . . . . . 8-1
Pages and page types . . . . . . . . . . . . . . . . . . . . . . . . 8-2
Data pages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
Row pages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
Extent pages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-7
Long varchar pages . . . . . . . . . . . . . . . . . . . . . . 8-11
Overflow pages. . . . . . . . . . . . . . . . . . . . . . . . . . 8-12
Index pages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-14
Control pages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-16
Database control block . . . . . . . . . . . . . . . . . . . . 8-16
Free extent list . . . . . . . . . . . . . . . . . . . . . . . . . . 8-17
Group free page list . . . . . . . . . . . . . . . . . . . . . . 8-17
Group extent map. . . . . . . . . . . . . . . . . . . . . . . . 8-17
Bitmap pages . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-17
Row count page . . . . . . . . . . . . . . . . . . . . . . . . . 8-18
Table data page . . . . . . . . . . . . . . . . . . . . . . . . . 8-18
Page allocation and garbage collection . . . . . . . . . . 8-18
Page allocation . . . . . . . . . . . . . . . . . . . . . . . . . . 8-18
Garbage collection . . . . . . . . . . . . . . . . . . . . . . . 8-20
9 B-Tree Indexes . . . . . . . . . . . . . . . . . . . . . . . . . 9-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
Advanced Topics Guide
Binary search trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-3
Multiway trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-4
B-Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-4
The order of a B-Tree . . . . . . . . . . . . . . . . . . . . . . 9-5
Sequencing within a node. . . . . . . . . . . . . . . . . . . 9-5
Balancing B-Trees . . . . . . . . . . . . . . . . . . . . . . . . 9-6
B-Tree operation costs . . . . . . . . . . . . . . . . . . . . . 9-9
B+-Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-10
Prefix B+-Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-12
SQLBase implementation . . . . . . . . . . . . . . . . . . . . . 9-12
Index node size. . . . . . . . . . . . . . . . . . . . . . . . . . 9-12
Performance considerations. . . . . . . . . . . . . . . . 9-13
Non-unique indexes . . . . . . . . . . . . . . . . . . . . . . 9-15
10 Hashed Indexes . . . . . . . . . . . . . . . . . . . . . . . 10-1
Hash table overview . . . . . . . . . . . . . . . . . . . . . . . . . 10-2
Row location in hashed tables . . . . . . . . . . . . . . 10-2
Performance benefits of hashed tables . . . . . . . 10-2
Potential pitfalls of hashing . . . . . . . . . . . . . . . . . 10-3
Hashing theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4
The hash function . . . . . . . . . . . . . . . . . . . . . . . . 10-4
Bucket hashing . . . . . . . . . . . . . . . . . . . . . . . . . . 10-5
Collisions and overflows . . . . . . . . . . . . . . . . . . . 10-5
Collision avoidance techniques . . . . . . . . . . . . . 10-6
Hashing disk space utilization and performance 10-7
SQLBase hashing implementation . . . . . . . . . . . . . 10-11
Location of table rows. . . . . . . . . . . . . . . . . . . . 10-11
Hash function . . . . . . . . . . . . . . . . . . . . . . . . . . 10-11
Specifying the packing density . . . . . . . . . . . . . 10-13
11 Estimating Database Size. . . . . . . . . . . . . 11-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-2
Spreadsheet usage. . . . . . . . . . . . . . . . . . . . . . . 11-3
Formulas for estimating database size . . . . . . . . . . . 11-3
Advanced Topics Guide
Calculation of column width . . . . . . . . . . . . . . . . 11-4
Table size calculations . . . . . . . . . . . . . . . . . . . . 11-6
BTree Index size calculations. . . . . . . . . . . . . . . 11-7
Estimation of overhead for views . . . . . . . . . . . . 11-8
Fixed amount of overhead . . . . . . . . . . . . . . . . . 11-8
Example of size estimation . . . . . . . . . . . . . . . . . . . 11-10
12 Result Sets
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-2
Characteristics of result sets . . . . . . . . . . . . . . . . . . . 12-2
Physical structure of result sets . . . . . . . . . . . . . 12-2
Logical contents of result sets . . . . . . . . . . . . . . 12-3
Population of result sets . . . . . . . . . . . . . . . . . . . 12-7
Flushing of result sets. . . . . . . . . . . . . . . . . . . . . 12-8
Processing result sets . . . . . . . . . . . . . . . . . . . . . . . . 12-9
Lock duration on result set rows. . . . . . . . . . . . 12-10
Concurrency and derived result sets . . . . . . . . 12-10
Concurrency and non-derived result sets. . . . . 12-11
13 Concurrency Control
. . . . . . . . . . . . . . . . . . 13-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-2
Transaction serialization . . . . . . . . . . . . . . . . . . . . . . 13-2
Lock types and scope . . . . . . . . . . . . . . . . . . . . . . . . 13-5
S-locks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-6
X-locks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-6
U-locks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-6
Lock compatibilities. . . . . . . . . . . . . . . . . . . . . . . 13-7
Locking level . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-7
Locking scope. . . . . . . . . . . . . . . . . . . . . . . . . . . 13-7
Lock implementation. . . . . . . . . . . . . . . . . . . . . . 13-9
Isolation levels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-10
Using isolation levels effectively . . . . . . . . . . . . 13-11
Selecting isolation levels. . . . . . . . . . . . . . . . . . 13-13
Advanced Topics Guide
14 Logging and Recovery . . . . . . . . . . . . . . . . 14-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-2
Failure types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-2
Transaction failures . . . . . . . . . . . . . . . . . . . . . . 14-2
System failures . . . . . . . . . . . . . . . . . . . . . . . . . . 14-3
Media failures . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-3
Recovery mechanisms . . . . . . . . . . . . . . . . . . . . . . . 14-4
The log file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-4
Backup and recovery . . . . . . . . . . . . . . . . . . . . 14-14
15 Introduction to Query Optimization . . . 15-1
Data access languages. . . . . . . . . . . . . . . . . . . . . . . 15-2
Procedural data access languages. . . . . . . . . . . 15-2
Declarative data access languages . . . . . . . . . . 15-3
Query optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 15-4
Syntax optimization. . . . . . . . . . . . . . . . . . . . . . . 15-4
Rule-based optimization . . . . . . . . . . . . . . . . . . . 15-5
Cost-based optimization . . . . . . . . . . . . . . . . . . . 15-7
Query optimization steps . . . . . . . . . . . . . . . . . . 15-7
16 Overview of the SQLBase Query
Optimizer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-2
Relational operations . . . . . . . . . . . . . . . . . . . . . . . . 16-3
Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-5
Selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-5
Join . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-6
Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-8
SQLBase physical operators. . . . . . . . . . . . . . . . . . . 16-9
Sort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-9
Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-9
Disk access operation. . . . . . . . . . . . . . . . . . . . 16-10
Join operations . . . . . . . . . . . . . . . . . . . . . . . . . 16-11
Imbedded operators . . . . . . . . . . . . . . . . . . . . . 16-14
Advanced Topics Guide
Query plan structure . . . . . . . . . . . . . . . . . . . . . . . . 16-15
Query trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-15
Building query trees . . . . . . . . . . . . . . . . . . . . . 16-16
Predicate logic conversion . . . . . . . . . . . . . . . . 16-17
17 SQLBase Optimizer Implementation . 17-1
Database statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 17-2
Table statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 17-2
Index statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 17-4
Selectivity factor . . . . . . . . . . . . . . . . . . . . . . . . . 17-6
Use of statistics by the optimizer. . . . . . . . . . . . . . . . 17-9
18 Working with the SQLBase Query
Optimizer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-2
Update statistics . . . . . . . . . . . . . . . . . . . . . . . . . 18-2
Optimal index set . . . . . . . . . . . . . . . . . . . . . . . . 18-2
Tuning a SELECT statement . . . . . . . . . . . . . . . . . . 18-3
Step 1: Update statistics . . . . . . . . . . . . . . . . . . . 18-3
Step 2: Simplify the SELECT command . . . . . . . 18-4
Step 3: Review the query plan . . . . . . . . . . . . . . 18-6
Step 4: Locate the bottleneck . . . . . . . . . . . . . . . 18-6
Step 5: Create single column indexes for the
bottlenecks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-7
Step 6: Create multi-column indexes . . . . . . . . . 18-8
Step 7: Drop all indexes not used by the
query plan. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-9
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Advanced Topics Guide
Advanced Topics Guide
This is an advanced reference manual for SQLBase. It contains detailed discussions
of various technical topics, including database design and SQLBase query optimizer.
Advanced Topics Guide
The Advanced Topics Guide is written for anyone using the advanced features of
SQLBase. This includes:
Application developers
Application developers build client applications that access databases using
frontend products like SQLWindows and the SQL/API.
Database administrators (DBAs)
Database administrators perform day-to-day operation and maintenance of
the DBMS and an organization’s databases. They install maintenance
releases, load data, control access, perform backup and recovery, and monitor
Database designers
Database designers, or senior database administrators, are responsible for the
logical and physical design of databases. Their responsibility is to ensure the
database meets the operational and performance requirements of the system.
This manual assumes you have:
Knowledge of relational databases and SQ
Familiarity with the material in the other SQLBase manuals.
What is in this manual
This manual is organized into the sections listed below, each of which is further
subdivided into a number of chapters. There is also an index.
Section Name
Database Design
Chapters 1-7 discuss logical and physical database design.
SQLBase Internals
Chapters 8-14 present several SQLBase internals topics.
SQLBase Query
Chapters 15-18 discuss the SQLBase query optimizer.
Advanced Topics Guide
Notation conventions
The table below show the notation conventions that this manual uses.
A developer who reads this manual.
The end-user of applications that you write.
bold type
Menu items, push buttons, and field names. Things that you select.
Keyboard keys that you press.
Courier 9
Builder or C language code example.
Program names and file names.
A plus sign between key names means to press and hold down the first
key while you press the second key.
These are numeric boolean constants defined internally in Builder:
Successful, on, set
Unsuccessful, off, clear
Advanced Topics Guide
Other helpful resources
Centura Books Online. The Centura document suite is available online. This
document collection lets you perform full-text indexed searches across the entire
document suite, navigate the table of contents using the expandable/collapsible
browser, or print any chapter. Open the collection by selecting the Centura Books
Online icon from the Start menu or by double-clicking on the launcher icon in the
program group.
Online Help. This is an extensive context-sensitive online help system. The online
help offers a quick way to find information on topics including menu items,
functions, messages, and objects.
World Wide Web. Centura Software’s World Wide Web site contains information
about Centura Software Corporation’s partners, products, sales, support, training, and
users. The URL is http://www.centurasoft.com.
To access Centura technical services on the Web, go to http:/www.centurasoft.com/
support. This section of our Web site is a valuable resource for customers with
technical support issues, and addresses a variety of topics and services, including
technical support case status, commonly asked questions, access to Centura’s Online
Newsgroups, links to Shareware tools, product bulletins, white papers, and
downloadable product updates.
For information on training, including course descriptions, class schedules, and
Certified Training Partners, go to http://www.centurasoft.com/training.
Other Publications. Depending on your requirements, you also use publications for
these products:
A Guide to the SQL Standard, C.J. Date
This book describes the SQL relational database language with an emphasis
on both the official standard version of SQL and on the use of SQL for
programmed (versus interactive) database access.
SQL & Its Applications, Raymond A. Lorie and Jean-Jacques Daudenarde,
Prentice Hall, 1991.
The Art of Computer Programming, D.E. Knuth, Addison-Wesley, 1973.
The Ubiquitous B-tree, D. Comer, ACM Computing Surveys, Vol. 11, No. 2
(June 1979).
Files and Databases, an introduction, P. Smith and G. Barnes, AddisonWesley, 1987.
Algorithms + data structures = programs, N. Wirt, Prentice-Hall, 1976.
Advanced Topics Guide
Relational Completeness of Data Base Sublanguages, E.F. Codd, in Data
Base Systems, Courant Computer Science Symposia Series, Vol. 6. PrenticeHall, 1972.
Concurrency Control and Recovery in Database Systems, Philip Bernstein,
Vassos Hadzilacos, and Nathan Goodman, Addison-Wesley, 1987.
An Introduction to Database Systems, C.J. Date, Addison-Wesley, 1990 (5th
Relational Database: Writings 1985-1989, C.J. Date, Addison-Wesley,
Relational Database: Writings 1989-1991, C.J. Date, Addison-Wesley,
The Relational Model for Database Management: Version 2, E.F. Codd,
Addison-Wesley, 1990.
Computer Data-Base Organization, James Martin, Prentice-Hall, 1977 (2nd
Query Evaluation Techniques for Large Databases, Goetz Graefe, ACM
Computing Surveys, Vol. 25, No. 2 (June 1993).
Join Processing in Relational Databases, Priti Mishra and Margarate Eich,
ACM Computing Surveys, Vol. 24, No. 1 (March 1992).
On Optimizing an SQL-like Nested Query, Won Kim, ACM Transaction on
Database Systems, Vol. 7, No. 3 (September 1982).
Join Processing in Database Systems with Large Main Memories, Leonard
Shapiro, ACM Transaction on Database Systems, Vol. 11, No. 3 (September
Send comments to...
Anyone reading this manual can contribute to it. If you have any comments or
suggestions, please send them to:
Technical Publications Department
Centura Software Corporation
975 Island Drive
Redwood Shores, CA 94065
or send email, with comments or suggestions to:
[email protected]
Advanced Topics Guide
Advanced Topics Guide
Chapter 1
This chapter introduces the following chapters on database design:
Chapter 2 - Conceptual Schema Creation
Chapter 3 - Internal Schema Creation
Chapter 4 - External Schema Creation
Chapter 5 - Transaction Definition
Chapter 6 - Refining the Physical Design
Chapter 7 - Physical Design Control
Advanced Topics Guide
ANSI/SPARC three schema architecture
The Database Design Section is organized around a database architecture that was
originated by the ANSI/X3/SPARC Study Group on Data Base Management
Systems. This group was established in 1972 by the Standards Planning and
Requirements Committee (SPARC) of ANSI/X3. ANSI is the American National
Standards Institute and X3 is this organization’s Committee on Computers and
Information Processing. The objectives of the Study Group were to determine the
areas, if any, of database technology for which standardization was appropriate, and
to produce a set of recommendations. The Study Group concluded that interfaces
were the only aspect of a database system that could possibly be suitable for
standardization, and so defined a generalized architecture or framework for a
database system and its interfaces.
This architecture includes three separate schemas (a schema is a model or formalism
to represent an abstract concept) of an organization’s data resource. Each schema
provides a perspective that serves a limited purpose and is optimized towards a
different set of objectives. The schemas included in the ANSI/SPARC three- schema
architecture are the conceptual, external, and internal.
The conceptual schema, sometimes referred to as the logical data model, is based on
the natural characteristics of the data without regard to how the data may be used or
physically stored. The conceptual schema models the data and the relationships
between data as they naturally exist. The assumption is that this perspective of an
organization’s data resource is the least likely to change over time and is therefore the
most stable. Likewise, it is assumed that how data is used and physically stored
changes frequently over the life of an organization. Consequently, the conceptual
schema is the primary or base schema within the ANSI/SPARC three schema
architecture to which the internal and external schemas are mapped.
The external schema is the user's views or filters of the database and specifically
considers how the data is used. The views are tailored to the user’s information
requirements and allow the user access to the required data in the simplest possible
form. Since the user’s data requirements change frequently, the external schema is
subject to constant revision.
The internal schema is a mapping of the conceptual model to the physical constructs
of the target database management system (DBMS). In performing this translation,
the database designer may introduce new physical objects or alter the conceptual
objects to achieve better performance or capacity and a more usable database. These
design changes can be made to exploit the hardware or software environment of an
implementation platform. Since these changes should not be visible at the conceptual
level, the actual meaning of the data does not change. Since the views of the data in
the external schema are also independent of this physical schema, users do not have
to deal with the physical objects.
Advanced Topics Guide
ANSI/SPARC three schema architecture
Through the three schemas and the mapping between them, the ANSI/SPARC three
schema architecture provides a level of indirection which insulates user needs from
database performance and capacity concerns. The major advantage of using this
architecture is that as the physical characteristics of the data and/or the target DBMS
changes over the life of the database, database designers can adjust the physical
constructs of the internal schema without affecting existing program modules.
Similarly, the user’s views of the data can be changed as their requirements change
without affecting the manner in which the data is physically stored. Without the three
schema (or similar) architecture, changes to the physical database can require costly
effort wasted on maintenance of existing program modules.
Although no commercial DBMS now provides full support for the ANSI/SPARC
three schema architecture, this architecture provides an excellent goal for the data
base administrator (DBA) to work towards. It also provides a good framework for
organizing database design activities.
Advanced Topics Guide
Structure of this section
In the remaining six chapters of this section we reveal a step by step path through the
database design process which specifies and optimizes each schema of the ANSI/
SPARC three schema architecture. This process is shown in Figure A.
Ch a pte r 2
C us t
O rd er
Conceptual Schema Creation
C ha p ter 3
Lo g ica l D a tab a se D esign
Internal Schema Creation
F irst-C u t D ata b as e
C ha p ter 4
External Schema Creation
U ser V iew s
Ch a pte r 5
Transaction Definition
Tra ns act ion D e fin itions
Ch a pte r 6
Refining the Physical Design
B etter D a tab a se
Ch a pte r 7
Physical Design Control
O p tim ized D a ta b a se
Figure 1A - Database design chapters and deliverables
Some DBAs do not like the work required to create and maintain the external schema
level of the ANSI/SPARC standard. Since the benefits of using this set of user views
are limited by the current state of the art in relational database theory (see chapter 4
for further details), they avoid this work until the payoff is more substantial. This is a
valid viewpoint, and organizations that wish to take this approach may omit the work
in Chapter 4, External Schema Creation. They can then code all of their transaction
data manipulation language (DML), as well as user queries, directly against the base
Advanced Topics Guide
Structure of this section
tables in the internal schema. The risk associated with doing this is that alterations
made to the physical design have a greater chance of affecting existing DML and
causing some re-coding to keep the DML valid. Note that creating the external
schema does not totally eliminate this risk.
Chapter 2 - Conceptual Schema Creation
Logical data modeling, the creation of the conceptual schema, is a (relatively) wellknown process within the database industry. Many books and training courses are
available on the subject. Consequently, this chapter does not attempt to present an
exhaustive treatment of this material. Rather, this chapter introduces you to the logical
database design process. Readers who are interested in a more detailed presentation
of the material are encouraged to read one of the many excellent books available on
the subject or attend a local training course.
In addition, the database industry has standardized on the “Entity-Relationship
Diagram,” or ERD, to represent the conceptual schema. Consequently, many
Computer Aided Software Engineering (CASE) vendors market tools for entity
relationship diagramming. You may find one of these tools beneficial in the creation
of your conceptual schema. Centura has established partnerships with several CASE
vendors to provide tighter degree of integration between the CASE tool and the
Centura software. At the time this manual was written, partnerships exist with LBMS,
Intersolv, Bachman, Popkin, Visual Systems and LogicWorks.
Chapter 3 - Internal Schema Creation
In this chapter we cover the tasks necessary to create the first-cut internal schema or
physical database design. The database constructed in this chapter is an initial
prototype only, representing the logical data requirements in their simplest form,
without any processing requirements. The first-cut internal schema is constructed by
applying some simple conversion rules to the entity relationship diagram. These rules
are discussed in detail in this chapter.
Also, many of the CASE tools discussed above automatically generate the first-cut
internal schema. You may find using a CASE tool which converts an entity
relationship diagram to SQLBase compatible SQL DDL (data definition language)
Chapter 4 - External Schema Creation
In this chapter we cover the tasks required to create the first user views needed for the
external schema. These views are mainly reflections of underlying tables without any
complex user requirements. We introduce you to the on-going process of view
administration, since the external schema evolves as new needs arise. We discuss
tasks of view administration and the underlying security requirements of views.
Advanced Topics Guide
Chapter 5 - Transaction Definition
We create transaction definitions at this point because they represent detailed and
critical processing requirements. The final database must meet these requirements in
order to be completed successfully. Completion of the tasks outlined in this chapter
provides the yardstick by which the rest of the database design process is judged.
Chapter 6 - Refining the Physical Design
In this chapter we re-visit the internal schema created in chapter three with the goal of
introducing physical database design optimization methods which improve its
performance and capacity. At this point in the database design process, general
guidelines and heuristics are applied to the first-cut physical database design to
produce a better physical database design, or internal schema. Over the years,
database designers have discovered many rules-of-thumb or heuristics to improve
overall database performance. Many of these common guidelines are detailed in this
Chapter 7 - Physical Design Control
In this chapter we conclude the database design effort by entering the next stage, the
iterative process of adjusting the internal schema to meet the objectives set in the
transaction definitions. Transactions that fail to meet processing requirements are
examined for possible improvements that target their particular database accesses.
Next we evaluate the effect on the other transactions that use the database. Our goal is
to find an acceptable trade-off between conflicting requirements of all of the
transactions which make up the complete system. This process is reiterative until all
critical transactions are performing effectively.
Advanced Topics Guide
Advanced Topics Guide
Chapter 2
Conceptual Schema
The conceptual schema is the primary or base schema within the ANSI/SPARC three
schema architecture and is sometimes referred to as the logical data model. It models
an organization’s data resource and the relationships between data as they naturally
exist without regard to how the data may be used or physically stored. Of the three
schemas, this is the least likely to radically change over time.
This chapter introduces the process of logical data modeling and several related
topics. In this chapter we:
Introduce the data administration function and responsibilities.
Cover logical database design.
Explain types of entity-relationship data models.
Explain the data normalization technique.
Advanced Topics Guide
Conceptual Schema Creation
Data administration
The data administration (DA) function plays a consulting role in the process of
logical database design. This contrasts with the traditional mission of the database
administration (DBA) function, which is mainly responsible for physical database
design and implementation tasks. The mission of data administration is to ensure that
an organization’s data resources are managed responsibly and recognized as an
important asset. A data administration department is responsible for formulating the
plans that guide the evolution of all corporate databases. By providing this master
plan, the DA function ensures that corporate computing systems will be compatible in
the future.
This compatibility guarantees that future system development will not adversely
affect existing systems and users in some way, and that information in separate
databases today can be combined in the future without undue cost or delay. Without
this assurance, a company can be exposed to financial and operational risk. This can
arise from either the disruption of mission critical computer systems or through the
company’s inability to meet future information requirements effectively.
Many corporations have not formally established a DA function. In the absence of
this formal mandate, it is important for those personnel responsible for operating and
maintaining the database environment to establish an agreement to “self-rule.” This
helps to avoid the pitfalls caused by a lack of standardization. Since database
professionals are often the first people to realize the benefit of establishing DA
practices, it behooves them to either attempt to persuade management to formally
establish the DA function, or to make it part of their regular procedures as best they
In this chapter we intend to provide a starting point for database administrators who
are working without formal DA functions and who have not yet adopted any DA
practices. There are many books and articles that treat these and other DA subjects in
great detail, and we encourage you to search out those subjects that interest you.
Standards creation and enforcement
One of the first tasks of data administration is the creation of numerous standards.
The policing of these standards is one of the major ongoing efforts of the DA
function. There are many opinions regarding the exact composition of these
standards, particularly naming standards. However, the most important aspect of
these standards is that they exist, are fully documented and distributed, and that they
are followed.
Advanced Topics Guide
Data administration
The following list represents some of the areas in which standards are generally
Standards for the abbreviations used as roots for object names are useful in
maintaining naming consistency.
Data element (or attribute) names
These standards benefit users who access the database using tools such as
Quest. These users must find the information they need in a consistent way.
When several people are creating new elements, variations of the same name
can be created. Good element and abbreviation standards can solve this
problem. Element naming standards should also address the capitalization
and word separators of multi-name elements.
Table names
The previous comments on data element names apply here as well. Names
must be simple and meaningful to users.
View names
Many sites want naming conventions that allow easy differentiation between
views and tables. Some sites like to highlight views that result in join
operations since these may be time consuming and have restrictions on update
Program names
Most installations try to indicate application ownership of programs by
mnemonics in the name. Mnemonics are short abbreviations, often two or
three letters, which contain enough of the original phrase to prompt an
association to the original. An example of a mnemonic is “AR” for accounts
receivable. Also, some sites like to be able to tell whether a program is readonly from the name alone.
File names
File names should bear some relationship to the database that owns the file.
User names
These standards have to do with who a person is and where that person can
be found, either by department or physical location.
The definition and utilization of standardized templates is valuable in
performing security authorizations for users based on the data requirements
of their job. This procedure is very useful when an audit is being performed.
Advanced Topics Guide
Conceptual Schema Creation
Database names
Standardizing database names simplifies operational control of the databases
for day-to-day use. The same applies to the following areas or items:
Server names
Workstation names
Printer names
Data dictionary
One of the key tools of data administration is a data dictionary or repository. This is a
special database which contains metadata (data about data such as element names,
meaning, data type, null characteristics, masks, and validations) describing the
business’s database environment. Within the dictionary is information about each
table, its columns, indexes, views, and relationships to other tables. Through the use
of the dictionary, users and programmers can find the data items, tables, or views they
need. Also, data administrators can monitor naming standards, locate data, and
eliminate inconsistencies in the ways various departments describe and use data.
Since the dictionary is the main tool needed for data administration, database
administrators who attempt to provide basic DA functions for their organizations can
design and implement a simple dictionary for their own use. A database describing
tables, views, and columns can be created and then populated with data extracted
from the SQLBase system catalog tables. These are included in every SQLBase
database, and have names which start with “SYSADM.SYS”, such as
“SYSADM.SYSTABLES.” By combining this information across all databases, the
DBA can locate and correct non-standard names, inconsistent key definitions, and
duplicate table or column definitions.
The Team Object Manager component of Centura Team Developer contains a
repository which you may use to store your metadata. Similarly, many CASE tools
include a data dictionary and you may find it beneficial to use the repository of the
CASE tool you have selected to perform database design tasks.
Enterprise-wide data planning
Corporations which conduct strategic management as an organization-wide activity
usually perform some type of strategic planning within the MIS division. Although
various strategic planning methods may be used, most techniques lead to the creation
of a strategic data plan. This plan serves as a guide for the development of the
corporation’s data resource throughout its development. Often, one of the outputs of
the strategic data plan is an enterprise-wide data model that can be used as a roadmap
by future database designers.
Advanced Topics Guide
Data administration
The format of an enterprise-wide data model is generally that of an entity-relationship
diagram (ERD). This diagram shows the major business entities in the organization
and the way they relate to each other. The entities should have descriptions attached,
but most attributes can be omitted when the modeling is done at this high level.
Include primary key attributes, along with their data types, in order to enforce
consistency among these critical attributes.
It is up to database designers working in organizations that have an enterprise-wide
data model to be familiar with the model and to keep database designs consistent with
it. While the enterprise model will be missing some low-level entity types and most of
the attributes, it should dictate the general structure of the major entities and their
relationships in any lower level design. If this is not true, then adjust the enterprise
model. Analyze and document any impact from making these changes.
Subject area databases
A database may be designed around either application systems or subject areas.
Research and experience has shown that building databases around subject areas
rather than application systems to be far superior.
For example, an organization could build a database around an accounts receivable
application, with its own customer entity, and then design another database around an
order entry application, also containing another customer entity. This results in both a
proliferation of databases as well as data duplication. While the customer entities in
the two databases just mentioned may differ in some ways, they probably have an
overlap of many attributes such as name and address. Keeping track of these two
entities’ common attributes is difficult and error prone. Most application systems
designed with their own database simply ignore the similar information in other
systems and make no effort to maintain consistency. Consequently, most
organizations with a formal data administration function design databases around
business subject areas rather than application systems.
Center the design around subject areas, such as product or customer, and your result
will be many different applications and user departments that can create, update, and
share information in a seamless fashion. In addition, you are likely to find that the
time necessary to develop application systems decreases significantly as the corporate
database is completed.
This is because a significant portion of most application development projects is spent
building the programs necessary to add, change, and delete persistent data. If
databases are designed around application systems, then a database must be built for
every application system and the programs to add, change, and delete persistent data
must be written for every application. On the other hand, if the database is designed
around subject areas with a high degree of data sharing, at some point in an
organization’s development cycle all subject areas are built, the database is complete,
and all the programs necessary to add, change, and delete persistent data are written.
Advanced Topics Guide
Conceptual Schema Creation
Once an organization reaches this stage, new applications can be developed rapidly
since no new programs must be written to add, change, and delete persistent data.
Organizations that have a formal DA function and have completed a strategic data
plan and an enterprise data model often move on to performing data modeling on
certain key business areas before starting development work on applications. These
business area data models are usually fully normalized (normalization is covered later
in this chapter) and complete, with specifications of all keys and attributes in each
entity included within the model. These data models serve as good sources of
material for a database design and can often be used as is, but two or more may have
to be merged to obtain the complete set of data required by one application.
Logical data modeling
The most rigorous technique now in use to perform the logical database design task is
logical data modeling. Using diagram symbols, a model (or picture) of the data
required to meet the objectives of the system is developed. The resulting model is
called an entity-relationship diagram (ERD).
Techniques are applied to this model to verify its accuracy, completeness, and
stability. The model can then be translated into a physical database design, possibly
for multiple implementation platforms, by performing a transformation process that
meets specific performance requirements as well as any special characteristics of the
chosen physical architecture.
The three most common styles of ERDs in use today are the Bachman diagram, the
Information Engineering diagram and the Chen diagram. Two are named after their
authors, while the other was specified by James Martin as a component of his
information engineering (IE) diagram set. Each style represents similar types of
objects, and examples of each follow. The most basic objects are the entity and the
relationship. Keys, attributes, and domains are not assigned diagram symbols, but are
important background objects for logical data modeling. Each of these concepts are
discussed in the following sections.
Entities and relationships
Entities are the basic information units of an ERD and are equivalent to tables in the
finished physical database design. Entities represent a person, place, concept, or event
in the real world about which information is needed. Examples of entities are
customers, orders, invoices, employees, warehouses, and so on. Entities are
represented by boxes in all three diagram styles.
Relationships are connections between one or more entities. These are represented
graphically by lines containing various symbols that convey characteristics of the
relationship. The majority of relationships involve only two entities and are called
Advanced Topics Guide
Logical data modeling
binary relationships. These are often broadly classified according to how many
occurrences of an entity type can appear on each side of the relationship, as shown in
Figure A.
One -to-one
One-to-ma ny
Man y-to-m a ny
Figure 2A - Types of relationships classified by maximum cardinality or relationship
An example of a simple ERD for each of the three diagramming styles follows. Each
diagram depicts the same model, a common example of a business’s basic order
processing data requirements.
The Bachman ERD style
ha s
pla ce d
ha s
w ith in
ha s
f or
Figure 2B - Bachman ERD of order processing data requirements
In the Bachman ERD, the CUSTOMER entity has a relationship with the ORDER
entity, as indicated by the line between the two entities. An important characteristic of
the relationship is its cardinality, which describes the allowable number of
occurrences of each entity type involved in the relationship. In the Bachman notation,
minimum cardinality (which can be either zero or one) is described through the use of
circles at the ends of the lines (a relationship with a minimum cardinality of zero is
often called optional, while those with a minimum of one are termed mandatory),
while maximum cardinality (which can be either one or many) is described through
Advanced Topics Guide
Conceptual Schema Creation
the use of arrows. In the case of minimum cardinality, the circle can either be hollow,
indicating zero, or filled in, indicating one. To describe maximum cardinality, the
absence of the arrow means a maximum of one, while the presence of the arrow
means a maximum of many (more than one).
In Bachman terminology, a relationship is called a partnership set, which is made up
of two partnerships. Each of these two partnerships is owned by one of the entities
involved in the relationship. In this example, CUSTOMER has a partnership with
ORDER, and ORDER also has a partnership with CUSTOMER. These two
partnerships together form the CUSTOMER-ORDER partnership set.
For CUSTOMER’s partnership with the ORDER entity, the minimum cardinality of
orders for a customer is specified by the hollow circle next to the CUSTOMER entity,
which is labelled “has.” The minimum cardinality is zero therefore this circle is
hollow. This means that a customer may exist without any orders. The maximum
cardinality of the ORDER entity from the perspective of the CUSTOMER is
described by the arrow on the ORDER side of the relationship, which specifies that a
customer may have many orders. If there were no arrow, then the maximum
cardinality would be one rather than many, and a customer could have only a single
order at any one time. The way to read this relationship in terms of the CUSTOMER
entity is to say, “A customer has from zero to many orders.”
From the ORDER perspective, the minimum cardinality of CUSTOMER is 1. This is
noted by the circle next to the ORDER entity which is labelled “placed by” and is
filled in. The maximum cardinality of the relationship is also one, since there is no
arrow on the CUSTOMER side of the relationship. This type of relationship, where
an entity has both minimum and maximum cardinality of one with another entity, is
often called a dependency. The ORDER entity is said to be dependent on
CUSTOMER, since it could not exist without a relationship to a single occurrence of
CUSTOMER. The reading from the ORDER entity’s perspective is, “An order is
placed by one, and only one, customer.”
In a similar manner, the other relationships read:
An order has 1 or many items.
An item is within 1, and only 1, order.
A product has from zero to many items.
An item is for 1, and only 1, product.
Advanced Topics Guide
Logical data modeling
The Information Engineering ERD style
Figure 2C - Information Engineering ERD of order processing data requirements
In the Information Engineering ERD, entities are also depicted by boxes and
relationships as lines. The significant difference from the Bachman ERD is the
graphic description of relationship cardinality. In this diagram, each entity’s minimum
and maximum cardinality in a relationship is described immediately adjacent to itself.
For instance, the CUSTOMER’s cardinality with respect to ORDER is shown in two
horizontal bars immediately below CUSTOMER. The symbol closest to an entity
describes the maximum cardinality and can be either a bar (indicating one), or a
crow’s foot (indicating many). Following the relationship line, the next symbol
describes the minimum cardinality and can be either a bar (indicating one), or a circle
(denoting zero). The meaning of each relationship in Figure C is the same as
described for Figure B.
Advanced Topics Guide
Conceptual Schema Creation
The Chen ERD style
Figure 2D - Chen ERD of order processing data requirements
In the Chen ERD, there is a more significant deviation from the previous examples.
Entities are still represented by boxes, and lines between entities still play a part in
representing relationships, but now there are also diamonds that appear on the
relationship lines which contain the name of the relationship. Also, relationships may
have attributes assigned to them directly. These are called information bearing
relationships. In the Chen diagram, attributes may appear directly on the diagram as
bubbles which contain the attribute’s name and are attached to the entity or
relationship to which they belong. Attributes have been omitted from this diagram in
order to remain uncluttered and consistent with prior diagrams. When this happens,
the relationship containing the attributes would be transformed into a table in the
physical database design in the same way that entities become physical tables.
The ORDER ITEM relationship is an example of this. In previous diagrams, ORDER
ITEM was an entity rather than a relationship. In the Chen diagram, the relationship
between ORDER and PRODUCT is indicated as having a maximum cardinality of
“many” on both sides (shown by the “M” and “N”), and the ORDER ITEM
relationship between them is information bearing and would appear as a table in the
resulting physical database design. The table that results from two entities having a
many-to-many relationship is said to resolve the relationship. The data contained in
the table is called junction data, because it is the information that is unique for each
occurrence of the relationship between the two entities. In this example, the junction
data would be attributes such as QUANTITY-ORDERED. Also, while the Chen
diagram indicates maximum cardinality by means of a number or symbol (the “1”,
“N”, and “M” in Figure D) next to the entity involved in a relation, minimum
cardinality is not depicted. In Figure D, “N” orders occur for each CUSTOMER
(where “N” indicates some number greater than 1), thereby depicting the customer to
order relationship as one-to-many. The “one” side of this relationship is indicated by
the “1” appearing next to the CUSTOMER entity.
Advanced Topics Guide
Logical data modeling
Attributes and domains
Attributes are simply characteristics of entities. They either describe or identify an
entity. Identifying attributes are also called keys, which are further explained in the
next section and are data items such as EMPLOYEE-NUMBER, CUSTOMERNUMBER and SSN (social security number). Examples of descriptive attributes are
Following the transformation of the logical design into a physical database, attributes
will become columns of tables.
Domains are the logical value sets that attributes possess. There is some variation on
the definition and use of domains in different diagram styles. Notably, the Bachman
ERD calls this definition of domains a dimension, which is further classified into
domains which then possess a particular display characteristic. The definition of
domain given here is the usage specified by E. F. Codd when he published the
relational data model. Attributes such as TOTAL-SALE-DOLLARS and PRODUCTCOST could share a domain called CURRENCY, for example. This is because the
possible values of these attributes is a discrete set of those real numbers that are
represented by the picture 999,999...,999.99. Likewise, the attributes of ORDERDATE and BIRTHDATE could share the DATE domain, in which 02/29/1992 is a set
member, but 02/29/1993 is not.
Domains provide a means to perform semantic validation of query statements,
particularly for join operations. When tables are joined on columns that are not
defined on the same domain, the results will not be meaningful. An example of this is
be specified for a data model since they are not at this point transformed into any
physical design object, their inclusion can provide a means of manually validating the
join criteria used in SQL statements. They are also useful in the physical design
process for determining physical data types of table columns.
Primary and foreign keys
The two most important types of keys for database design are primary and foreign
keys. These provide the means by which SQLBase supports referential integrity. For a
logical data model to be complete, every entity must have a primary key identified.
This primary key can consist of either a single attribute, or may be composed of two
or more attributes in which case it is called a concatenated or composite key. Some
diagramming techniques allow mandatory relationships to be specified as an
alternative to attributes when defining primary keys. When this is done the effect is
for the primary key of the entity on the other side of the relationship to be inherited as
a component of a concatenated key in the dependent entity.
Advanced Topics Guide
Conceptual Schema Creation
The two critical characteristics of a primary key are uniqueness and non-redundancy.
Uniqueness refers to the fact that the value of a primary key for a given occurrence of
an entity must not be duplicated in any other occurrence of the same entity type. The
physical design implication is that a unique index must be able to be placed on a
primary key successfully. The non-redundancy characteristic of a primary key, which
applies to concatenated keys only, means that each of the attributes which comprise
the primary key must be essential to its uniqueness, so that if any attribute was
removed, then the key would no longer be unique. This last rule prevents extraneous
information from being placed into keys. These rules for forming primary keys are
important in the normalization process.
Leg end :
P rim ary K ey
F oreig n K ey
Figure 2E - Example of mapping between primary and foreign keys. Dependent entity
ORDER_ITEM inherits part of its composite key from ORDER. The ORDER-NO
attribute also serves as the foreign key to ORDER.
Another rule many organizations follow in formulating primary keys is to avoid
building concatenated keys through stringing together many coded attributes. As an
example, consider a part number where codes are present that indicate the product
family, weight, color, primary vendor, and voltage of the part. Some users like this
type of key because they can tell a great deal about the part merely by looking at the
key. The drawback, however, is that when these attributes change (such as a new
vendor), the primary key has to change. When the primary key changes, any
references to it as a foreign key in related entities must change as well. All of these
changes to data rows consume processing time and also contribute to the need to
perform more frequent database reorganizations to prevent performance degradation.
Since keys often have indexes placed on them, all these indexes will become
increasingly disorganized as well.
Advanced Topics Guide
Foreign keys must be created in an entity type to indicate the specific occurrence of
another entity with which this entity has a relationship. The attributes of the foreign
key must match the primary key of the related entity exactly. The foreign key appears
in the entity that is on the many side of a one-to-many relationship, or may appear on
either side of a one-to-one relationship. Many logical data modeling diagram
techniques do not explicitly name foreign key attributes, since their existence can be
assumed through the relationships that the entity participates in. When the
relationship is many-to-many, an intermediate table will contain the foreign keys of
both entities involved in the relation. See Figure E for the relational notation of the
primary to foreign key mapping for the data model used in the previous examples.
The normalization process is a logical design technique that is especially well suited
for use with SQLBase (or for that matter, any DBMS which similarly adheres to the
relational model of data) because of its capability to directly implement the results
without any translation. Performing normalization results in reduced data redundancy,
protects against update and deletion anomalies, and provides a more flexible, longer
lasting database design. The resulting database will be easier to maintain and will not
require major, high-cost structural changes as often as unnormalized designs. This is
due to the inherent stability of basing the data structures on the natural organization of
the basic business information flowing through the corporation rather than structuring
it with only a limited user or application perspective.
While the processes that a business performs on its data may change often, the data
items that are operated on by those processes can be structured to reflect the nature of
the company’s business. This means that the application processes, embodied in
program logic, can be more quickly changed to keep pace with company needs. The
underlying database remains the same, reflecting the basic information inherent in the
environment. Attaining this level of stability is the goal of data normalization.
One of the fundamental requirements of a designer who wishes to perform
normalization is to know the meaning of the data intimately and to have a thorough
understanding of its use in the real world of business. The way that a table’s attributes
interact with and depend on each other is the basis for normalization. The designer
performing the normalization process must be knowledgeable of the data at this
detailed level in order to successfully complete the process without undue risk of error
arising from misunderstandings or assumptions.
Advanced Topics Guide
Conceptual Schema Creation
Normal forms
The normal form is a specific level of normalization achieved by a set of tables
comprising a database design. Each successive normal form represents a more stable
and controlled stage of design. Higher levels of normal forms always require the
design to be in each lower level of normalization as a prerequisite. As an example,
second normal form requires the design to be in first normal form, third normal form
requires both first and second forms, and so on.
D escription
Un norm alize d D ata
First No rm al Form
R ecords w ith no repea tin g
g ro ups
S econd No rm al Fo rm
A ll non key data it em s fully
func tiona lly dep endent on
p rim ary key
Third No rm a l Form
A ll non key data it em s
indep end ent o f ea ch eac h
Decom pose all da ta structures into fla t,
tw o -dim e nsio na l tab les.
For tables w ith com p osite prima ry keys,
en sure th at all o ther a ttrib utes are dep en dent
on th e w h ole key. S p lit ta ble s a pa rt as
required to ach ieve th is.
Rem ove all transitive d epe nd en cie s, splittin g
th e table apa rt as req uired. No n on key a ttribute
sho uld rep re sen t a fact abou t a not her no nkey
at trib ute .
Figure 2F - Summary of normal forms, rules, and steps needed to achieve the next
There are a number of normal forms defined and work is continuing so that more
forms may appear in the future. Most current texts on database design will address
normal forms one through five, although only the first three are needed for most
tables. For this reason, only the first three forms are summarized in Figure F. Consult
more detailed database design texts for a more in-depth discussion of additional
normal forms.
Advanced Topics Guide
First normal form
Figure 2G - Conversion of an unnormalized relation, STUDENT, to three separate
relations in first normal form. Primary keys are indicated by the attributes which
Conversion of an unnormalized relation called STUDENT into first normal form is
shown in Figure G. This conversion involves breaking the table apart, with each table
having its own primary key identified. Any repeating elements would also be removed
from the relations at this point, so that each resulting table is “flat”, meaning that each
column has a unique name and no columns consist of multiple elements (such as an
array). This simple step places these data items into first normal form.
Advanced Topics Guide
Conceptual Schema Creation
Second normal form
Figure 2H - Further normalization of the previous example, where the CERTIFICATE
relation violates second normal form. The small arrows indicate which key attributes
each non-key attribute is dependent on. The other relations in the previous example
are already in second normal form.
Further transformations are required to bring the data structure from Figure G into
second normal form. As Figure H shows, the CERTIFICATE relation violates second
normal form, and must be broken apart. The non-key attributes of Certificate_Name
and Certificate_Hours_Reqd are only dependent on part of the primary key,
Certificate_Code. The table is therefore transformed so that the Certificate_Code is
the entire primary key, with the two attributes which depend on it remaining in the
table. The Certificate_Date_Earned attribute, which depended on the entire key of the
original CERTIFICATE relation, is placed into a new relation,
STUDENT_CERTIFICATE, along with the original composite key. The entire set of
relations is now in second normal form.
Advanced Topics Guide
Third normal form
Figure 2I - Further normalizing our example to third normal form. The student
relation violated third normal form due to non-key attributes which are dependent on
other non-key attributes, as shown by the arrows. These attributes are broken out into
new relations, the original relation as a foreign key to the new relation. This foreign
key attribute is MAJOR_DEPT-CODE in the STUDENT relation.
One remaining transformation must be done on our data from Figure H in order to get
all the relations into third normal form. As shown in Figure I, the STUDENT relation
violates third normal form, and must be transformed into two tables. The two non-key
attributes Major_Dept_Name and Major_Dept_Location depend on another non-key
attribute, Major_Dept_Code. Even though the are currently in the STUDENT table,
they do not depend on the Student_Number primary key directly, but are dependent
through the Major_Dept_Code, which is dependent on the primary key. This is called
a transitive dependency, and is removed by creating a new table with a primary key of
the transitive attribute in the original relation. This transitive attribute then becomes a
foreign key to the new table. The attributes in the original relation that were
transitively dependent are now placed in the new table, where they are fully
dependent on the primary key only. The MAJOR relation in Figure I is formed
through this process. The data originally described in Figure G is now entirely in third
normal form.
Advanced Topics Guide
Conceptual Schema Creation
Now that we have looked at the benefits associated with data normalization, a few
words about denormalization are in order. Denormalization is the process of taking a
database design out of third normal form (or some higher form) for performance
reasons. One may denormalize by either redundantly storing prime attributes, or by
adding aggregate attributes to the database.
Some database designers routinely denormalize during logical database design.
However, the only reason to denormalize is to achieve greater performance. In
addition, because of the added complexities of a denormalized database (refer to
chapter 7 for further details), denormalization should only occur as a last resort after
all other means of increasing performance have failed. Consequently,
denormalization should only take place during physical design control. The logical
design should still be maintained in (at least) third normal form in order to facilitate
clear thinking about its structure and meaning.
Advanced Topics Guide
Advanced Topics Guide
Chapter 3
Internal Schema Creation
In this chapter we cover the tasks necessary to create the first-cut physical database
design (internal schema). The database constructed in this chapter is an initial
prototype only. The entities, relationships and the other logical constructs of the
conceptual schema are mapped to SQLBase physical constructs without regard to
performance or capacity issues. The objective of this step is to produce SQLBase
compliant SQL DDL (data definition language) which represents the initial physical
database design. The first-cut internal schema is constructed by applying several
simple conversion rules to the entity relationship diagram. These rules are discussed
in detail in this chapter.
Many of the CASE tools discussed in chapter 1 automatically generate the first-cut
internal schema. You may find using a CASE tool which converts an entity
relationship diagram to SQLBase compatible SQL DDL beneficial.
Advanced Topics Guide
Internal Schema Creation
Building the initial internal schema
Following completion of the logical database design, the designer must convert the
entity relationship diagram (ERD) to SQL data definition language (DDL) which
specifies the desired physical database to SQLBase. This transformation may be done
either manually by the designer, or in an automated fashion through the facilities of a
CASE tool.
The following steps are required to perform the conversion:
Create the base tables. These are the basic building blocks of the database,
and are derived from the entities contained within the ERD. Each possesses
a number of characteristics which must be considered by the designer:
The table-wide options specified on the CREATE TABLE statement,
such as owner, name, and PCTFREE.
The list of columns contained within the table, which is derived from the
attributes of the original ERD entity.
The specific data type associated with each column. These may have to
be determined now, or may have been specified as part of the ERD
domain specification.
Create any required junction tables needed to resolve many-to-many
relationships in the ERD.
Add referential integrity (RI) specification to the DDL. This implements the
relationships specified in the ERD, and requires the following steps:
Identify the primary key of each table within the CREATE TABLE
Build unique indexes for each primary key of each table.
Build ALTER TABLE statements which identify the foreign keys
contained within child tables.
Creating tables
The first step in building the ANSI/SPARC internal schema is to identify the
relational tables which will be managed by SQLBase. The primary input to this task
is a completed ERD containing entities and relationships. On completion, a set of
CREATE TABLE skeleton statements may be coded that will serve as the starting
point for the addition of columns and other details.
A base table will be created for each entity on the ERD. Base tables are the major
information-bearing objects within a relational database, and contrast with views in
Advanced Topics Guide
Creating tables
that no projection or selection operations may restrict the visibility of any columns or
For each base table identified, define a long identifier which uniquely names the table
within this database. This name should follow site standards for base table names
while also being as descriptive as possible about the business information contained
within the table. Also, use an explicit authorization-id instead of allowing the
authorization-id to default to the user name of the current user. Select an
authorization-id that meets your site standards, and can serve as a consistent owner
for all application objects within the database.
When you have finished, verify that the number of tables matches the count of entity
boxes in the E-R diagram.
Adding columns to tables
The columns needed by the tables you have created are present in the ERD as
attributes. These attributes need to be transformed into column specifications in the
middle of the CREATE TABLE statements you will code. The attribute names
themselves must be transformed into column names, and must adhere to the SQLBase
long identifier restrictions. This is a good time to check that the root portions of the
column names (that part between underscore delimiters) conform to the standardized
list of abbreviations. New abbreviations should be evaluated for acceptability and
added to the list. Column names which are not clear and consistent may cause
confusion for programmers and users throughout the life of the application.
Data typing for columns
After identifying the columns, data types need to be added. This task will be greatly
simplified if the ERD contains domains and the attributes have all been given domain
assignments. First convert each domain into an appropriate data type, if necessary.
Some ERDs may have domains which are already specified in terms of RDBMS data
types as implemented by SQLBase. These require no further transformation. When
performing this conversion, consider the following factors:
Internally, SQLBase stores all columns as variable length fields. Therefore,
choosing between some data types becomes a matter of documentation,
which may be useful for range validation purposes. As an example, defining
a classification code column such as LEDGER_SUBCODE in a general
ledger application as CHAR(3) documents the fact that values must always
consist of three characters (perhaps falling into the range “10A” to “99Z”)
more clearly than defining the data type as VARCHAR(3). Each of these two
data type definitions results in an identical internal implementation in
SQLBase, however.
Advanced Topics Guide
Internal Schema Creation
If the usage of a numeric data item requires it to be of a fixed width, use the
DECIMAL data type with the required precision and an appropriate scale. If
an integer is desired, use DECIMAL with a scale of zero. This helps to
communicate external formatting requirements more clearly.
Avoid the use of INTEGER and SMALLINT data types in general. Use these
only for internal control fields such as cycle counters. Formatting these fields
for external display use can be problematic due to the indeterminate display
Avoid using LONG VARCHAR unless absolutely required. Each LONG
VARCHAR column must be stored on a separate page, apart from the rest of
the row to which it belongs. Also, each of these columns will occupy at least
an entire page by itself, regardless of how much free space on the page may
be wasted. Therefore, using LONG VARCHAR increases both the number
of I/O operations to read a row, as well as the amount of space required to
store the data. In addition, some @ functions cannot be used with LONG
VARCHAR columns.
Avoid using CHAR or VARCHAR for data that will always be numeric.
Using a numeric data type will ensure that the field will always be populated
with valid numeric data.
Avoid the use of the REAL (or FLOAT or DOUBLE PRECISION) data type
except for situations calling for scientific measurement information
containing a wide range of values. The exponent-mantissa format of these
data types lacks the exact precision associated with other numeric data types,
and also causes floating point operations to be performed for their
manipulation. On most CPUs, floating point arithmetic incurs greater cycle
times than integer operations.
Use the DATE and TIME formats when expressing chronological data.
Use DATETIME (or TIMESTAMP) for control purposes. Many designers
include a TIMESTAMP field for tracking row creation and updates.
Null usage
When defining data types for columns, you will also need to specify the column’s
ability to assume the null value. Note that null is a special value that indicates either
“unknown” or “not applicable”. Users and programmers not accustomed to the use of
nulls may have problems remembering to include host variable declarations and tests
for nulls in columns that allow them. Nulls can also cause problems when they are in
columns that are used for joining tables together. When the join is performed, any
rows that contain nulls in the join column of either table will not be returned to the
result set. This “missing data” may be disconcerting to a user of the query, who may
have expected to see all of the rows of at least one of the tables. Of course, an outer
join specification (using the “+” operator) may be used to force all rows of one of the
Advanced Topics Guide
Creating tables
tables to be returned, but this requirement could be easily overlooked. Consider the
following factors when deciding whether to allow a null value in a column:
Columns participating in primary keys should always be defined with NOT
NULL, since they are required to be populated with unique values for each
Foreign keys should generally be defined with NOT NULL. Whenever a
child table is dependent on a parent, the foreign key to the parent table cannot
be null, since the existence of a child occurrence without a parent would
violate the dependency rule of the relationship. Only foreign keys to tables
with optional relationships can be considered as candidates for the use of
nulls, to signify that no relationship exists. Also, foreign keys defined with a
delete rule of SET NULL must be defined as nullable.
Use NOT NULL WITH DEFAULT for columns with DATE, TIME, or
DATETIME data types to allow current date and time data to be stored in the
field automatically. This is particularly useful for row creation timestamps.
Allow nulls for those columns that will actually have an anticipated need for
the unknown or inapplicable meaning of the null value. The null value works
especially well in numeric fields that may be subjected to aggregation uses
such as SUM or AVERAGE.
Use NOT NULL WITH DEFAULT for all columns not meeting one of the
above criteria.
Example 1
We will carry forward the example ERD used in the previous chapter to demonstrate
diagramming styles to show the transformation from external to internal schema. The
following SQL statements represent the current state of the design at this point. Note
that a site standard requires base table names to be suffixed with a “_T” to facilitate
easy identification.
Advanced Topics Guide
Internal Schema Creation
Resolving many-to-many relationships
If the ERD contains any many-to-many relationships, these cannot be directly
implemented in the relational model without creating an intermediate table. This
process is called resolving the relationship. The relationships affected are those that
exist between two tables that have a maximum cardinality of many on both sides of
the relationship. These relationships are transformed into two one-to-many
relationships with a new table which serves as a placeholder for the purpose of
representing occurrences of related parent tables. Figure A, below, contains an
Advanced Topics Guide
Resolving many-to-many relationships
example of resolving a many-to-many relationship between CUSTOMER and
Figure 3A - An example of resolving a many-to-many relationship through the
creation of a junction table. The junction table CUST_SLSMAN contains the primary
key of each parent and is also dependent on both. Examples of other columns that
could be contained within the junction table would be
CUSTOMER and SALESMAN tables remain unchanged. Columns which are primary
keys are underlined, while foreign keys columns are in italics.
Resolving the relationship
Create a new table, called a junction table, between the two existing tables. Often,
designers will name this table as a conjunction of the names of the two tables
involved in the relationship. Figure A shows an example of this by creating a
CUST_SLSMAN junction table between the CUSTOMER and SALESMAN
Add columns to the new table. The required columns are the primary keys of both
tables in the relationship. Be sure to include all columns of any composite primary
Define primary and foreign keys for the junction table. The primary key will be a
composite key made up of the concatenation of the primary keys of both tables in
the relationship. The foreign keys will be each parent table’s half of the primary
key, which references back to that parent table. This completes the transformation
Advanced Topics Guide
Internal Schema Creation
of the single many-to-many relationship into two one-to-many relationships with
an intermediate junction table.
Now that the many-to-many relationship has been resolved, take a moment to think
about other attributes that may exist within the relationship. This information, called
junction data, can be added as columns to the new table that resolved the relationship.
There are often data items that will appear in these tables as the project progresses
and user requirements become better known. Two examples of such columns are
given in Figure A.
Adding referential integrity
Referential integrity is the verification that links between table rows are valid at all
times. These links are the physical implementation of the relationships modeled in
the Entity-Relationship diagram during logical design. The links are implemented
through common values contained in primary and foreign keys in the related rows.
These rows may exist in the same table, or in separate tables. The table containing the
primary key is said to be the parent table in the relationship, and the table containing
the foreign key is termed the child table. (Although these terms “parent” and “child”
are widely accepted in relational terminology, they connote a hierarchical, dependent
relationship between the tables. Of course, a relationship implemented through
primary and foreign keys in SQLBase, or any other RDBMS, could be of several
types other than a hierarchy.) It is important to remember that these terms are only
meaningful in the context of a single relationship, since a child table in one
relationship could be a parent table in another relationship, or vice-versa.
By defining a table’s primary and foreign keys when you create the table, SQLBase is
able to enforce the validity of relationships within the DBMS itself without requiring
any special coding conventions by the user. For instance, without system-enforced
referential integrity, an order entry program would have to retrieve a row from the
customer table (corresponding to the customer number entered by the operator) in
order to verify that the customer actually existed. With referential integrity, the child
order table contains the customer number as a foreign key to the parent customer
table and the program no longer needs to check for the existence of the customer.
When the order row is inserted, the DBMS will realize that the referential integrity
definition requires it to perform a read of the customer table to verify that a row with
a primary key that matches the contents of the foreign key in the order record exists.
If this row is found in the customer table, then the order row can be inserted
successfully. Otherwise the insert will fail and an appropriate SQL error code will be
returned to the program to indicate that a referential integrity error occurred.
Sometimes concerns are raised about the possible performance penalties that can be
incurred through the use of system defined and enforced referential integrity. While
there may be some small minority of cases where user enforced referential integrity
Advanced Topics Guide
Adding referential integrity
could be performed more efficiently, in the vast majority of cases, the DBMS itself is
able to maintain a higher level of consistency with a lower cost. The two most
important factors contributing to this are its ability to perform integrity checks for all
database accesses, and its ability to always use the most efficient access technique
available to perform the referential checks.
Selection of appropriate referential integrity constraints requires an understanding of
the business rules governing the way applications can delete and update data. Use of
this knowledge allows the database designer to select the appropriate delete rules
from among the three options available in SQLBase. These options are:
Restrict, which disallows deletion of parent rows if any child rows exist. This
is the default, and is generally the safest option.
Set Null, which sets the foreign key values of child rows to null when parent
rows are deleted.
Cascade, which deletes child rows when parent rows are deleted.
There are also implications of delete rules resulting from the transitive relationships
between three or more tables that can cause restrictions to be placed on the designer’s
selection. Furthermore, some DML coding options can be constrained through the use
of referential integrity. The designer must consider all these factors when selecting
appropriate referential integrity constraints for the database.
Defining keys
Name the columns comprising the primary key of each table in the PRIMARY KEY
clause of the CREATE TABLE statement. Since the primary key is also required to be
unique, the next step is required before the primary key definition is complete.
Following each table’s CREATE TABLE statement, code a CREATE UNIQUE
INDEX statement that builds a unique index on the primary key columns of the table.
Include only those columns that are in the primary key, and index them in the same
order as they are specified in the PRIMARY KEY definition clause of the CREATE
TABLE statement. Consider using some indicator (such as a preceding, or following,
“XPK”) in the index name to signify that this is a primary key index, since
inadvertently dropping this index would have detrimental effects on the desired
referential integrity for the database. It is important to remember that both the
PRIMARY KEY clause of the CREATE/ALTER TABLE statement and the CREATE
UNIQUE INDEX are required before SQLBase will consider the primary key
definition for the table to be complete.
At the end of the DDL file (following all CREATE TABLE/INDEX statements), code
an ALTER TABLE statement that adds the foreign key definitions needed for each
child table in the database. Adding foreign keys through this ALTER technique
guarantees that each table’s primary key definition is completed (with its associated
index) prior to being referenced in a foreign key clause by some other table.
Advanced Topics Guide
Internal Schema Creation
Example 2
Continuing from example 1, we now incorporate the required referential integrity
definitions to define the relationships in the ERD. These modifications include the
PRIMARY KEY clause of the CREATE TABLE statement, the CREATE UNIQUE
index statement for each primary key, and finally, the ALTER TABLE statements
needed to define the foreign keys.
Advanced Topics Guide
Adding referential integrity
Advanced Topics Guide
Advanced Topics Guide
Chapter 4
External Schema
In this chapter we cover the creation of the initial external schema. This requires
coding the data definition language necessary to build views on the underlying tables
of the internal schema, called base tables. It also means establishing the procedures
needed to allow the external schema to evolve as requirements for new user views of
the data become known. Security mechanisms may also be put into place to insure
that all data access is accomplished through the external schema, rather than by direct
access to base tables. In this way, the base tables are allowed to change while isolating
the user views from some of the impact of these modifications.
Advanced Topics Guide
External Schema Creation
Using views as the external schema
In the ANSI/SPARC three schema architecture, the external schema is used to isolate
the user’s (or program’s) data requirements from the physical implementation. This
isolation is intended to benefit users by allowing them to be unaware of the physical
aspects of the database structure, which were possibly decided for reasons of physical
DBMS and hardware features and limitations. This external schema is also isolated
from the conceptual schema so that the user’s data requirements are not impacted by
any esoteric modeling processes (such as normalization) or organization-wide data
policies that may seem to have unnatural effects on their own data. In other words, the
goal of the external schema is to allow data to be presented to the user in as natural
and simple a way as possible, and to allow for their data requirements to change
dynamically as the business needs evolve. The facility that can be used to accomplish
the goal of the external schema in RDBMSs, such as SQLBase, is the SQL view.
However, with the current SQL standard used by most RDBMSs, including
SQLBase, the support for this external schema is limited. The current SQL standard
fails to differentiate between conceptual and physical objects, and currently fails to
provide an external view mechanism that transcends the ability of the physical design
to impact user views. However, just because these limitations currently exist does not
mean that they will not be solved in the future as the SQL standard evolves. The
relational view mechanism can currently be used to isolate users’ data requirements
from many types of physical changes. Theorists and researchers are currently
working to extend the ability of views to be isolated from physical database
dependencies. Much of the current limitations manifest themselves as certain view
DML limitations, the most notable being the inability to process any update
commands when a view accesses multiple tables, such as in a join.
Examples of physical schema level database changes that can be ‘hidden’ by views
include the following:
Object owners changes. Changing the owner of views can have an impact,
Column names changes. The view definitions can map the old column names
to the new names, thereby making column name changes invisible to users’
Table names changes. These can be handled the same as column names
mentioned above.
Table column additions, since these new columns need not be added to
existing views.
Table column re-ordering. Views can explicitly order their columns
independent of the underlying table’s column ordering.
Advanced Topics Guide
Using views as the external schema
Index creation, modification, or removal. Of course, these are also ‘invisible’
to base table users. Note that this does not apply to primary key indexes,
where changes or deletions can impact referential integrity, which is not an
external schema impact but rather has an impact on the conceptual schema.
The combination of two or more tables into one. This assumes that the
primary keys of the two tables are the same.
The splitting of a table into two or more tables. This can be hidden if the table
is split so that no view has to access both new tables to access its required
This list could expand significantly as new releases of SQLBase become available. By
enforcing the use of the external schema through the view mechanism now, an
organization will be positioned to take advantage of present and future benefits
accruing to the three schema architecture without incurring costly source code
The primary costs incurred by the extensive use of views fall into two administrative
areas. The impact on the database administration area is the overhead of creating and
managing the many views that will exist. The second impact is on the task of security
administration, since more authorizations will need to be managed to ensure that all
database access is through views as opposed to base tables.
The view administration impact can be minimized by adopting good conventions for
managing the views. Views should be named in a meaningful manner and cataloged
according to the business data they supply. This will allow DBAs to maximize the
reuse of views by facilitating the identification of appropriate existing views as an
alternative to user requests to create new views. In many cases, programmers and
users may be able to locate a view on their own that provides the data they need
without involvement from the DBA.
The major pitfalls to avoid in view administration are the temptation to delegate the
responsibility of creating views too widely and the use of a very few broad views to
fulfill all access requests. The first pitfall comes about when the DBA grants the
ability to create views to nearly everyone in an attempt to eliminate the work of
creating views to DBA. While a few trusted individuals can be given this ability (if
they are carefully monitored), when many users have this ability, they may succumb
to the practice of simply creating a new view for every new access requirement they
have, without first searching for an existing view to satisfy their requirement.
The second pitfall occurs when the DBA avoids the workload of view creation by
satisfying every view request with the name of a view that mirrors the base table
definition of each table needed by the user. While this action is sometimes
appropriate, the DBA must be willing to create new views when no existing view
definition comes close to satisfying the new request. The DBA needs to find an
appropriate balance between these extremes by comparing the number of columns
Advanced Topics Guide
External Schema Creation
and rows that may be returned from a view with the result set that would exactly
satisfy the user’s request. For example, a rule of thumb could be set that specifies a
desired result set being within 75% of the columns and rows returned from a view as
The security administration work involved in enforcing the use of views can be
minimized as well. Techniques such as adopting automated procedures for the
creation and execution of the data control language (DCL, the SQL GRANT and
REVOKE statements) can greatly improve the consistency of security authorizations
as well as ease the workload. Making routine use of these procedures at key points in
the application development life cycle lowers the requirement for unscheduled, ondemand security administration.
Creating the initial external schema
To create the initial external schema for a new database, start by creating a view that
mirrors each base table in the internal schema. These views can be used by
programmers when they are coding DML statements to initially populate the
database, since they will have no update restrictions placed on them. These views are
also likely to be widely used by application programs that perform insert and update
operations. The name of each view should reflect the fact that it is a view of the
underlying base table, such as PART for the mirror view of the base table PART_T
(where the _T signifies a base table).
This mirror view should explicitly name each column in the base table, as opposed to
using the ‘SELECT *’ form to implicitly reference table columns, and contain no
WHERE clause that restricts row retrieval. If ‘SELECT *’ is used to implicitly
reference table columns, then application programs may require modifications when
columns are added or dropped from the base table. This is because application
programs must bind attributes returned from SQL statements to application program
variables and these bind clauses depend on a strict number and ordering of attributes
returned from SQL statements.
Additional views can be created for programs which perform database queries when
the data requirements of the program are fully documented. Also, views for ad-hoc
database users can also be created when their data requirements have been adequately
While these CREATE VIEW statements build the initial external schema view of the
database, this task is best viewed as a continual evolutionary process. As additional
information requirements become known, additional views will be added to the
external schema to satisfy these requirements in an efficient and flexible manner. In
order to facilitate the administration of the external schema’s DDL, these views
should be kept together in a subdirectory where they can be conveniently accessed
and re-executed, should the need arise.
Advanced Topics Guide
No programs or queries should be written which access the base tables, either by the
DBAs or anyone else. In order to enforce the use of the external schema as the only
means of accessing the database, do not grant authorization on the base tables to
anyone other than DBAs. Even the DBAs should exercise restraint in their use of the
base tables, using them only for utilities which exclude views from being named. The
views themselves can either be granted to PUBLIC or to individual users, depending
on how restrictive the organization’s security policies are.
While these security restrictions may seem excessive to some, it is only through their
usage that the goal of using the external schema to isolate user data views from
internal implementation can be approached and eventually achieved.
Carrying forward the example of the previous two chapters, we will now create the
mirror views of the base tables:
Advanced Topics Guide
External Schema Creation
Now we will create two views for known user requirements. The first view is for the
aged open order line inquiry and the second view is for the daily cancelled order
Advanced Topics Guide
Advanced Topics Guide
Advanced Topics Guide
Chapter 5
Transaction Definition
This chapter:
Explains the purpose of well-defined transactions:
Database design benefits
Data integrity considerations
Covers the detailed information required for each transaction:
Identifying information
Categorical information
Volume information
Performance requirements
Data requirements
Advanced Topics Guide
Transaction Definition
Benefits of defining transactions
Transaction definitions are required for proper physical database design and to ensure
data integrity.
Physical database design
The primary goal of physical database design is to ensure that the database provides
the required level of performance. Typically, database performance is measured in
terms of transaction performance. So, the objective of physical database design is to
ensure every database transaction meets or exceeds its performance requirements.
The process of physical database design involves mapping the logical data structures
(entities and relationships in an ER diagram) to the physical constructs of SQLBase.
For example, entities in the logical database model become one or more clustered or
non-clustered SQLBase tables. The design criteria which guides this mapping are
transaction performance requirements and volume statistics. The deciding factor for
the designer choosing between the various physical constructs that may be used is
their effect on transaction performance. For any particular construct, the effect will
probably enhance the performance of some transactions while degrading the
performance of others. So, physical database design can be viewed as an optimization
problem. The goal is to select the SQLBase physical constructs so that every
transaction meets or exceeds its performance requirements.
Hence, well-defined transactions are a prerequisite for physical database design.
Without this information, the database designer has no rule to guide in the selection
of physical database constructs. For example, the database designer will be unable to
determine whether a clustered or non-clustered table will result in better
performance. A similar argument can be made for index selection.
Data integrity
Database transactions usually involve multiple database accesses. For example,
consider the banking transaction of withdrawing $1000 from a savings account and
depositing the funds into a checking account. This transaction entails two separate
database operations: subtracting $1000 from the savings table and adding $1000 to
the checking table. If either one of these database operations successfully completes
but the other does not, then the account balances would be wrong. Both of these
operations must complete successfully or neither of them should complete
Thus, a database transaction is a logical unit of work which causes the database to
advance from one consistent state to another. In other words, a transaction moves the
database from one state which does not violate the data integrity requirements of the
database to another state which does not violate the data integrity requirements of the
Advanced Topics Guide
Defining transactions
database. In addition, a database transaction is a grouping of logically related
database operations so that all of the database accesses should complete successfully
or none of them should be applied to the database. Otherwise, if only some of the
database accesses complete successfully, the database is left in an inconsistent state.
So, in order to ensure the integrity of the database, it is necessary to rigorously define
the database transactions. This chapter describes how to define transactions, and
identifies the information which is typically included in transaction definitions.
Defining transactions
Transaction definitions can take many forms. Some organizations may possess an
upper CASE tool with a transaction definition facility as part of a data repository.
Other organizations may choose to define transactions using paper-based or
automated forms. Regardless of the media, all good transaction definitions include
several key ingredients. Each of these is discussed below.
Transaction name, number, and description
The first step in defining transactions is to uniquely identify each database
transaction. This is accomplished by assigning a name and unique identifier to each
In addition, a short narrative should be composed, referred to as the transaction
description, which describes the transaction in business terms:
The description should specify what the transaction accomplishes for the user
rather than how it is accomplished.
The description should be understandable by users (it should not include
technical jargon).
The purpose of transaction names and descriptions is to provide users and MIS staff
with a means for distinguishing one transaction from another.
Transaction Number: 001
Transaction Name: Transfer Funds
Transaction Description: The Transfer Funds transaction verifies
that sufficient funds exist in the account to be debited. If sufficient
funds exist, the transaction debits the account by the specified
amount and credits the other account for the same amount.
In OnLine Transaction Processing (OLTP) systems, the database transactions are
known ahead of time. For these types of systems, a one-to-one relationship will exist
between the transaction definitions and the actual transactions submitted to SQLBase.
Advanced Topics Guide
Transaction Definition
In Decision Support Systems (DSS), transactions are not known ahead of time. These
systems are characterized by ad hoc and exception reporting. Therefore, it is not
possible to describe each and every transaction. With these systems, the transaction
definitions are illustrative of the actual transactions which will be submitted to
SQLBase once the system is implemented. DSS systems require the database
designer to predict the type of transactions which the users will likely run.
Transaction type and complexity
Frequently, transaction definitions also categorize transactions in terms of type and
complexity. For example, transactions may be defined as either “Batch” or “Online”
and given a complexity rating of either “High,” “Medium,” or “Low.” This
information is useful for producing listings or totals of like transactions. For example,
management may require an estimate of the time required to complete physical
database design. To accomplish this, one might need to know how many low,
medium, and high complex transactions a system has. The database for a system with
a large percentage of high complexity transactions would take longer to design than a
system with mostly low complexity transactions.
The complexity rating is a matter of judgement, but should be based on the degree of
difficulty the database designer is likely to have meeting the performance
requirements for the transaction (see below). A transaction of high complexity has at
least two of the following characteristics:
Contains many SQL statements (more than 10).
Contains WHERE clauses with many predicates (for example, three or more
table joins or subqueries).
The SQL statements affect many rows (> 100).
On the other hand, a low complexity transaction has the following characteristics:
Contains few SQL statements (three or fewer).
Contains WHERE clauses with only one or two predicates.
The SQL statements affect few rows (< 25).
Transaction volumes
Transaction definitions must also include volume information. This typically includes
the average and peak frequency of each transaction. For example, a branch bank
might estimate that the funds transfer transaction described in the introduction of this
chapter will occur 50 times an hour with peak loads reaching 65 transactions per
The transaction volume statistics are extremely important for physical database
design; consequently, care should be taken to be as accurate as possible. For example,
Advanced Topics Guide
Defining transactions
the database designer will handle a transaction which occurs 1000 times per hour
quite differently from one that will occur only once per hour.
If the new computer system is replacing an existing system, it may be possible to
derive transaction volumes from the existing system. On the other hand, new systems
rarely duplicate the transaction logic of their predecessors so these differences should
be considered.
Typically, the hardest situation in which to deduce accurate transaction volumes is
one in which no automated or manual predecessors exist. These situations occur when
organizations enter new lines of business. In this case, volume statistics may be no
more than educated guesses.
Transaction performance requirements
The transaction definitions should also document the performance requirements of
each transaction. As mentioned in the introduction, the transaction performance
requirements are the basis of physical database design. During physical database
design, the database designer chooses the physical constructs from among the various
options provided by SQLBase, such that all transactions meet or exceed their
transaction performance requirements. Without specifying performance requirements,
the database designer has no means by which to evaluate whether or not the database
performance is acceptable.
Some organizations prefer to specify exact performance requirements for each
transaction. The performance requirements typically take the form of number of
execution seconds. For example, one requirement might state that the transfer funds
transaction must complete in three seconds or less.
Other organizations find it easier to define performance requirements in terms of
transaction types and complexity. For example, an organization might require:
high complexity, online transactions to execute in ten seconds or less.
medium complexity, online transactions to execute in six seconds or less.
low complexity, online transactions to execute in three seconds or less.
high complexity, batch transactions to execute in 60 minutes or less.
medium complexity, batch transactions to execute in 30 minutes or less.
low complexity, batch transactions to execute in 15 minutes or less.
Relative priority
Transaction definitions also should rank each transaction in terms of relative priority,
which describes how important to the business one transaction is compared to all
other transactions. Physical database design ensures that each transaction meets or
exceeds its performance requirements. Consequently, the transactions must be
Advanced Topics Guide
Transaction Definition
evaluated one at a time. The relative priority of each transaction provides the
sequence in which transactions are evaluated. This ensures that the most important
transactions receive the most attention.
In addition, during the process of physical database design, the database designer will
tailor the physical database design so that one given transaction executes faster.
Making one transaction go faster frequently negatively affects the performance of
other transactions. Relative priority ranking also provides the basis for any
transaction arbitration scheme.
The ranking can take many forms. For example, some organizations assign a number
from one (high) to ten (low) to each transaction.
Transaction SQL statements
Each transaction definition must also specify the SQL statements which perform the
necessary database operations. These SQL statements are either UPDATE, INSERT,
DELETE or SELECT commands, which are known collectively as SQL DML (data
manipulation language). The SQL statements are required during physical database
design to forecast the execution times of the transactions. In addition, a detailed
understanding of the database design and significant experience with SQL is required
to accurately code many SQL statements. Therefore, it is sometimes preferable for
the database designer to provide the SQL statements for the application programmers
rather than allowing them to code the SQL themselves. However, the database
designer and the application programmers must coordinate their efforts closely to
ensure that the SQL statements support the program's data requirements.
The transaction definitions should also include, for each SQL statement, a brief (nontechnical) narrative explaining:
What the command accomplishes
Why it is required (if it is not obvious)
Number of rows in the database affected by the command
The number of rows affected means either the number of rows returned (in the case of
a SELECT command), or the number of rows inserted, updated, or deleted. The
performance of any SQL command is dependent on the number of database rows
affected. For example, it would take longer to read 10,000 customer rows than one
customer row. Hence, the number of records affected by each SQL command is an
important element of physical database design.
Once coded and described, each SQL command should be validated for accuracy. To
accomplish this, it will be necessary to construct a small test database. Many users
find it useful to build a SQLTalk script containing all the transactions’ SQL
commands for this purpose.
Advanced Topics Guide
Defining transactions
Also keep in mind that the DML should be coded against the external, rather than the
internal, schema. The initial external schema created in chapter 4 will probably
contain many of the views needed to support the transactions. Other views may need
to be created in order to meet the tailored needs of some transactions. If the programs
are coded directly against the base tables contained in the internal schema, there will
be a greater risk of impacting them during the physical design control phase of the
database design.
SQL Statement
select amt from account
where acct_nbr = :DebitAcct
update account
set amt = amt - :TransferAmt
The table below shows how the SQL for the funds transfer
transaction might appear in a transaction definition.
This SQL command retrieves the balance of the debit
account. If it is less than the transfer amount, the program
should report an “insufficient funds” error to the teller.
Also, the SQL statement should use either the RR
isolation level or the SELECT FOR UPDATE command
to ensure that no other transactions withdraw funds from
the debit account while this transaction is in progress.
This SQL command decreases the funds in the debit
account by the transfer amount.
where acct_nbr = :DebitAcct
update account
set amt = amt +:TransferAmt
This SQL command increases the funds in the credit
account by the transfer amount.
where acct_nbr = :CreditAcct
Advanced Topics Guide
Advanced Topics Guide
Chapter 6
Refining the Physical
In this chapter we re-visit the internal schema created in chapter 3 with the goal of
introducing physical database design optimization methods which improve its
performance and capacity. At this point in the database design process, general
guidelines and heuristics are applied to the first-cut physical database design to
produce a better physical database design, or internal schema. Over the years,
database designers have discovered many rules-of-thumb or heuristics to improve
overall database performance. These guidelines are detailed in this chapter.
Advanced Topics Guide
Refining the Physical Design
In this chapter we present rules which may be used to refine the first cut physical
database design. These rules are typically applied in the following manner:
The database designer recognizes a scenario in the database which is often
the subject of a particular design technique.
The designer then applies that technique to the database on the assumption
that the benefits will be both substantial and required.
As such, these rules do not offer specifically proven benefits for every case, but
represent common database design rules that are usually incorporated into physical
designs by experienced DBAs. You should carefully evaluate every database as a
unique application, where some unusual characteristics may cause the assumptions
on which these rules are based to be violated. For these cases, as well as for further
refinement of any design, the more rigorous predictive techniques presented in
Chapter 7 may be applied.
These guidelines detailed in this chapter are discussed under the following general
Splitting tables into two or more tables
Clustering of table rows
Database partitioning
Freespace calculation
Splitting tables
One of the most common denormalization techniques that may be applied to a
physical database design is the vertical or horizontal splitting apart of tables. Vertical
splitting is the process of moving some of a table’s columns to a new table that has a
primary key identical to that of the original table. Horizontal splitting is the moving
of some of the rows of one table to another table that has a structure identical to the
first. The resulting tables can be managed in ways that tailor each to its particular
performance requirements. The basic motivation for performing a split is that either
the original table has some performance problem that must be solved, or one or more
subsets of the original table has significantly different performance requirements that
cannot be met in combination.
Advanced Topics Guide
Splitting tables
Vertically splitting large rows
The most common reason for vertically splitting apart a table due to performance
problems is that the row length of the original table is too large. There are two
common situations when the row length is too large: when the row length is larger
than the database page size and when a clustered hashed index is used.
When the row length is larger than the database page size
Some entities in the ERD may contain a large number of attributes or a few unusually
long attributes. When these entities are converted to a SQLBase table, the size of a
row can be actually larger than the usable database page size. In this case, when a row
is inserted into the table, SQLBase will allocate one or more extent pages on which to
store the row (see chapter 8, Database Pages). Consequently, multiple physical I/O
will be required to retrieve the row, one for the base page and one for each extent page
(worst case scenario). These multiple I/O operations will degrade performance. To
resolve this situation, the table can be vertically split in order to bring the row size
down to a more manageable size.
SQLBase’s usable page size is approximately 938 bytes. If the average row length for
a table is greater than the usable page size and optimum performance is required when
accessing the table, the database designer should split the table into two or more
tables. The method used to perform the split is simple, since the split is simply the
physical manifestation of performing a relational projection on the table. This means
that some of the columns in the original table will be moved to a new table, while the
remaining columns will reduce the row length of the table. The new tables that are
created using this technique have not changed any inter-column relationships in any
way (which means each column should still be fully dependent on the entire primary
key), therefore each new table should have an exact copy of the original table’s
primary key as its own primary key.
Determining which columns to place in each table is the most critical decision
affecting the successful application of this technique. This is because the split can
only result in a benefit to transactions that can satisfy their data requirements by
accessing only one of the resulting tables. Any transactions that must perform a join
to reassemble the original row will fail to benefit from the higher blocking factor due
to the additional I/O operations required to read the new table. To determine which
columns to place in each table, the database designer should analyze the access
patterns of all high priority transactions and place together in the same table those
columns which are accessed together the most frequently.
Advanced Topics Guide
Refining the Physical Design
When using a clustered hashed index
When a table’s data is clustered through the use of a clustered hashed index and the
row is too large, it fails to achieve the desired page clustering objective.
Consequently, the number of rows that will be retrieved for each physical I/O
operation may no longer be sufficient to provide adequate transaction performance.
This long row length is of concern when transactions commonly access a number of
rows with each execution, and the rows are physically clustered through a clustered
hashed index on the key that defines this group of rows.
The concept of page clustering or rows per page is the same as that of the blocking
factor for a file in general, where the number of rows that will fit on a page is
described by:
NumberOfRows = ----------------------rowsize
The purpose of splitting a table apart is to attempt to achieve a number of rows per
page that is at least as great as the number of rows that will be retrieved together by
common transactions. Since the size of the page is fixed by the implementation of
SQLBase on each hardware platform, the only way to influence the number of rows
per page is by adjusting the size of the row. Since there is usually no such thing as too
many rows per page, the direction this adjustment takes is to reduce the row size in an
attempt to increase the number of rows on a page. After eliminating any spurious
column definitions and reducing the maximum width of VARCHAR and CHAR
columns to a minimum, the only remaining technique is to split the table apart.
Horizontally splitting tables
As defined above, horizontal splitting of a table is the moving of some of the rows of
one table to another table that has a structure identical to the first. Horizontal splitting
occurs most often to isolate current data from historical data so that the performance
of accessing current data is increased.
For example, frequently organizations must store several years of order information
to produce comparison and historical reports. If an organization creates several
hundred thousand orders per year, an order table which contains five years of order
information might exceed one million rows. Such a large table would slow the access
to current year’s orders, which are probably the most frequently accessed. In this
situation, the database designer might choose to horizontally split the order table(s)
into two tables, one for the current year orders and another for all other orders.
Advanced Topics Guide
Splitting tables
Tailoring performance requirements
Whenever some rows or columns of a table have significantly different performance
requirements than the remainder, splitting the table apart is a viable option. This is
worthwhile if there are tangible benefits to be accrued from allowing the portion of
the table with the lowest requirements to accede to that level.
A good example of vertically splitting a table for this purpose is a LONG VARCHAR
column that actually contains a binary CAD/CAM drawing which is historical in
nature and rarely accessed by the users. Since this is a lot of data that has a much
lower usage rate than the remainder of the table, it could possibly be placed in some
lower cost storage medium, such as optical storage, in the future when appropriate
hardware and software become available. (Future releases of SQLBase may allow the
database designer to map specific tables to specific physical files and devices.) You
can accomplish this by splitting the table apart, with the LONG VARCHAR in a new
table that has the same primary key. This technique allows the LONG VARCHAR to
be stored economically (in the future) without affecting the other columns in the table.
If the column remains with the original table and future releases of SQLBase do add
this functionality, all existing programs which access the table will have to be
modified before the column may be mapped to these types of “near-line” storage.
Referential integrity considerations
When splitting a single table into two or more tables, the primary keys of each table
should be identical if the table was correctly normalized prior to the split. The new
tables’ relationships to other tables in the database, therefore, could also be the same
as the original table. However, the DBA should consider the effect of the delete rules
in the original table’s relationships when deciding how to define the referential
integrity constraints for the new tables. In some cases, it may be simpler not to define
referential constraints at all, since the actual row deletions from the new tables will
have to be maintained in parallel with the original table through programmatic means
An alternative technique is to make the new tables dependents of the original table,
with one-to-one relationships back to the original table. When this technique is used,
the new tables will have a foreign key back to the original table that is an exact match
to the primary key of the table. If the delete rule for this foreign key is cascade, then
SQLBase will automatically delete rows of the new tables whenever a corresponding
row of the original table is deleted. This eliminates the programmatic work required
to keep the tables in parallel for deletions, although insertions still have to be made
Advanced Topics Guide
Refining the Physical Design
Splitting apart a table
Determine which new tables will contain which columns of the original table.
Create the new tables with primary keys identical to the original table. This means
that each primary key column must be duplicated in each new table. Remember
to build the primary key index on each new table as well.
If you would like SQLBase to handle referential integrity constraints for the new
tables in the same manner as it does the original table as regarding its participation
as a child in relationships, then the foreign key columns for each parent must also
be duplicated in each new table. Furthermore, the new tables must contain the
FOREIGN KEY clauses that identify the parent table of each relationship.
If you would like to set up the new tables as dependents of the original table,
specify the primary key as a foreign key back to the original table. If you wish
SQLBase to handle row deletions from the new tables in parallel with the original
table, specify CASCADE as the delete rule for this foreign key. No other foreign
keys need to be defined in this alternative.
If you would like SQLBase to handle referential integrity constraints for the new
tables in the same manner as it does the original table regarding its participation
as a parent in relationships, then you must add a new FOREIGN KEY clause to
each child table of the original table that identifies this new table as an additional
If you have selected the alternative of keeping the new tables as dependents of the
original table in step 3, then the new tables should not participate as parents in any
other relationships. The original table should keep all of the relationships in which
it is the parent.
The new tables may be populated by means of an INSERT statement with a
subselect referencing the columns in the original table that need to be moved.
Drop the unneeded columns from the original table through an ALTER TABLE
Change any views that access the original table to access whichever tables contain
the columns needed by the view. If a view has to be changed from a single table
access to a join, determine whether any programs using the view are performing
updates. These programs will need source code changes to eliminate the attempt
to update through a view.
Advanced Topics Guide
Clustering of table rows
Clustering of table rows
Controlling the physical storage location of rows within a table using SQLBase is
accomplished through the use of a clustered hashed index. Using this indexing
technique, which is actually not a physical index at all but a method of storing the
underlying subject table, can enhance some aspects of database performance in
particular cases. There are two major reasons why this technique should be applied to
a table: one allows for faster access to single table rows while the other
accommodates the collocation of table rows that are often accessed together.
Direct random access of rows
Whenever fast direct access to a single row of a SQLBase table is critical, and a
unique key exists to identify that row, you should consider a clustered hashed index. A
unique clustered hashed index on the table column that is specified with an equal
predicate in the WHERE clause of a SELECT statement will cause the SQLBase
optimizer to access that row with a hash technique that will, in most cases, allow
retrieval of the page containing the row with a single read operation. This provides the
best possible access time for the row within the SQLBase system, exceeding the
capability of B+ index trees.
This technique would be appropriate, for example, when a high volume transaction
has to access a large operational master, such as an account table in a bank’s teller
system. Since the teller knows the customer account number, the program can access
the account information directly with a single I/O, eliminating not only the need to
traverse a large index tree, but also the I/O and locking overhead associated with that
task. Since many tellers are performing similar transactions simultaneously, the
overall load on the DBMS and disk subsystem is decreased through the use of a
clustered hashed index, while the response time for the transactions is enhanced.
In order for a table to be a good candidate for this technique, two conditions must
A good primary key must exist that can be fully specified for each access
request. Since a clustered hashed index does not provide any partial key
search capability (such as NAME LIKE ‘JO%’), the key must be one that can
be reliably specified in full. Furthermore, this key must also provide
satisfactory mapping to the physical database pages through the SQLBase
hash function. Since the hash function provided with SQLBase has been
designed to give good results for the majority of keys, this will probably be
no problem. Chapter 10 provides the details on how to verify the mapping of
the hash function for a given key.
The number of rows in the table must be fairly constant and predictable. This
is because the clustered hashed index CREATE statement must include an
estimate of the total number of rows that the table will contain. If the table
Advanced Topics Guide
Refining the Physical Design
size varies significantly over time, the benefits of creating the clustered
hashed index may be lost. A possible technique to get around this problem,
though, would be to drop and recreate the table and index often, with the row
occurrence estimate updated each time the index is recreated. This method
also requires an unload and load of the table as well, and for large tables may
be too time consuming.
Clustering rows together
When a fairly small number of rows of a single table are often accessed as a group,
the clustered hashed indexing technique can allow for these rows to be located
physically adjacent to each other. If the rows will all fit on one page, then a query
requesting all of the rows will be satisfied with a single I/O operation by SQLBase.
This allows for the maximum database performance to be realized for a commonly
accessed group of rows.
This technique is sometimes called foreign key clustering because it is often a foreign
key within the table that defines the group of data that should be collocated. That is,
when the value of a foreign key is the same for a group of records, SQLBase attempts
to store this group of records in proximity to each other. This foreign key becomes the
subject key of the non-unique clustered hashed index that is placed on the table.
A good example of this situation is in the familiar order entry system. The ORDER
and ORDER_ITEM tables in this system are often accessed together, and the
database designer wishes to make this access as quick as possible, by defining a
clustered hashed index on the ORDER_ITEM table, with the index column being the
ORDER_NO column. This column is the ORDER_ITEM table’s foreign key back to
the ORDER table. Since transactions first access the ORDER table to get the header
row for the order, they can then access the ORDER_ITEM table with a WHERE
ORDER_ITEM.ORDER_NO = ORDER.ORDER_NO clause. This will cause
SQLBase to use the clustered hashed index to locate this group of rows within the
ORDER_ITEM table. These rows will (hopefully) all be on one page and can be
retrieved with a single read operation.
An important qualifying requirement for the use of this method is the physical size of
the cluster that will be formed by grouping with the selected foreign key. If this
cluster size is too large, then the grouping will be spread out over many pages anyway
and the potential benefits of this technique will not be realized. Also, if the number of
rows within the grouping is large, then the chances of encountering collisions within
the hashing algorithm increase, causing an increase in overflow page allocations,
resulting in degraded performance.
Advanced Topics Guide
Clustering of table rows
In order to determine the average size of the foreign key grouping in terms of rows,
the following formula may be used:
grpsize = -------------fkeys
The numerator, which is the total number of rows in the table, may be determined
through estimation or through the use of the following query if the table already
SELECT COUNT(*) FROM tablename;
The denominator, which is the total number of unique foreign key values, may be
estimated based on the number of rows in the parent table where this foreign key is
the primary key. Alternatively, the following query may be run if the child table
already exists:
SELECT COUNT(DISTINCT ForeignKeyColumnName) FROM tablename;
This calculation yields the average number of rows for each foreign key group
If the resulting group size is very large (a good rule of thumb is greater than 20), then
the table is probably not a good candidate for clustering. The group size may be offset
somewhat by the row length, however. The smaller the row length, the greater the
group size that can be tolerated. This is because the true measure of how quickly
SQLBase will be able to retrieve all the rows in a foreign key grouping is determined
by how many pages are required to store these rows. The following formula can be
applied to determine the number of pages which will typically be required (remember
to calculate the correct page size for the implementation platform on which you will
run SQLBase):
( grpsize × rowle ngth )
pages = --------------------------------------------------------pagesize
If this formula yields a page spread of more than one, overflow will become a
problem and this table is probably not a very good candidate for a hashed clustered
index based on this foreign key.
Creating a clustered hashed table
If the table already exists, unload the data and drop the table.
Execute the CREATE TABLE command to create the table, along with the
CREATE UNIQUE INDEX statement that completes the primary key definition.
You may omit the CREATE UNIQUE INDEX only if the clustered hashed index
you will be creating is both unique and based solely on the primary key.
Code and execute the CREATE INDEX commands which builds the clustered
hashed index. Make sure that the SIZE clause is included, and that your estimate
for the number of rows includes enough space for growth that will occur in
Advanced Topics Guide
Refining the Physical Design
between database reorganizations. Also, consider allowing for free space of
between 20 and 30 percent which will achieve the minimization of hash collisions
(multiply your final estimate by 1.2 or 1.3 to do this).
Recreate any foreign keys in other tables which reference the primary key in this
Reload the data unloaded in step 1, if any.
Following this procedure, the entire space needed for the table will have been preallocated within the database.
An index in a relational database is a data structure that allows rapid access of table
rows based on the value of data in one or more columns of those rows. (For a detailed
explanation of SQLBase’s implementation of BTree indexes, see chapter 9). There
are only two reasons to build an index on a relational table. The first is to ensure the
uniqueness of the values being indexed, which is why all primary key columns must
have a unique index defined on them. The other reason for creating indexes (and the
only reason to create a nonunique index) is to enhance the performance of the
The drawback of indexes is the amount of resources that they consume. Every index
(other than a clustered hashed index) requires disk space for the storage of the index
tree structure. Each of these indexes also requires CPU cycles in order to be created
and maintained. In particular, indexes consume processing time whenever a
transaction performs an INSERT or DELETE operation against an indexed table, or
performs an UPDATE of a table row column that participates in one or more indexes.
Exactly how much overhead is created depends on the characteristics of the index and
subject table, but the overhead is not trivial in any case.
The challenge for the database designer is to create indexes only for those cases
where the benefits realized from the access time improvements outweigh the
penalties suffered from the overhead of building and maintaining the index. While
some scenarios can be identified as potentially good (or bad) candidates for an index,
no irrevocable ‘rules of thumb’ exist. For each index that a database designer creates,
the costs and benefits should be measured in order to justify the index’s continued
existence. If the designer discovers that an index is no longer worth its cost, then it
should be dropped.
As complex as this cost/benefit analysis is, it is further complicated by the fact that
the benefits are not actually within the control of the DBA. In SQLBase (as with any
DBMS implementation using the SQL standard), the decisions controlling how the
database is accessed are not made by the programmer, the user, or the DBA. Rather,
they are made by a software module within the DBMS called the optimizer.
Advanced Topics Guide
Therefore, the indexes which are finally selected by the DBA are subsequently
reviewed through a pre-determined statistical analysis by the optimizer, where they
may be found devoid of value and not used at all. Actually, the optimizer will select
some indexes for some DML statements, other indexes for other statements, and no
indexes for still other statements. For this reason, an accurate evaluation of the
benefits side of the index problem requires a good understanding of the optimizer and
how to monitor the decisions made by it.
The purpose of this section is to help the designer select possible indexes that are
worth considering within a physical database design. These must then be evaluated
for their effectiveness with particular DML statements through an analysis of the
decisions made by the optimizer.
Column cardinality and selectivity factor
The cardinality of a table column is the number of discrete values that the column
assumes across all rows of the table. As an example, the cardinality of the
EMPLOYEE-SEX column in the EMPLOYEE table is two, assuming that ‘M’ and
‘F’ are the only acceptable values. Another example is the EMPLOYEE-NUMBER
column in the same table, which is the primary key of the table and has a unique index
placed on it as the system requires. The cardinality of this column is equal to the
number of rows in the table, since for each row the column must assume a unique
value. These two examples illustrate the extreme cases of cardinality that may be
The reason cardinality is important to index design is that the cardinality of an index’s
subject columns determines the number of unique entries that must be contained
within the index. In the above example concerning the EMPLOYEE-SEX column, an
index placed on this single column would only have two unique entries in it, one with
a symbolic key of ‘M’ and one with a symbolic key of ‘F’. Each of these entries
would be repeated many times, since between the two entries all the table’s rows
would be accounted for. If there were 100,000 rows in the EMPLOYEE table, then
each index entry would be repeated 50,000 times if the distribution of employees
were equal between the sexes. This reiteration of the same symbolic key causes the
index to be unattractive to the SQLBase optimizer and can also have a significant
impact on the efficiency of index operations.
Determining the cardinality of a potential index key is simple if the data is already
loaded into the subject table. Use the following query to find the cardinality:
SELECT COUNT (DISTINCT key1||key2||...||keyn) FROM tablename;
In this query, key1 through keyn represent the potential index columns. The count
returned will be the exact cardinality that the index would have if created on
tablename at the current point in time.
Advanced Topics Guide
Refining the Physical Design
The way that the optimizer considers the effect of cardinality is through the use of a
statistic called the selectivity factor. The selectivity factor of an index is simply the
inverse of the combined cardinality of all index columns:
SelectivityFactor = -----------------------------Ca rdina lity
The selectivity factor is critical because this is the criteria used by the SQLBase
optimizer to decide whether an index is acceptable for use in the access path that is
built to satisfy the data requirements of a given DML statement. The optimizer uses a
formula that combines the selectivity factor of the chosen index, along with
information about the index’s width and depth, in order to compute an estimate of the
amount of I/O needed for a particular access path. The lower the selectivity factor, the
lower the resulting I/O estimate will be, which increases the possibility of the index
being selected for use.
These characteristics of columns and indexes will be used during the descriptions of
good and poor index candidates which follows.
Composite indexes
When an index is created on two or more columns, it is called a composite index.
When creating composite indexes, consider the affect that the order of columns
within the index has on performance. The best performance is achieved by placing the
column that usually has the most restrictive selection criteria used against it in the
high order index position. If you do not know which column meets this criteria, then
place the column with the highest cardinality first. Following these guidelines will
result in the lowest possible selectivity factor to be achieved by the index. Of course,
if the intended use of the index is to provide a generic search capability, then the
columns which will be always fully specified must be placed first in the index or the
searching capability will be lost.
One good reason to use composite indexes is to achieve index only access. This is
when all of the data which is required by a query can be found within the index itself.
A good example is a lookup table that finds tax rates for various county codes. If a
composite index is placed on the county code followed by the tax rate, then a query
that specifies an equality predicate on the county code within the WHERE clause, and
requests retrieval of the associated tax rate may be completed with an access to the
index entry without any access to the page containing the actual row.
A table is used by transactions to perform a lookup of county tax rates. Since
this is done often, the DBA wishes to allow the operation to be performed
with a minimum number of I/O operations. This is the SQL that is the most
efficient in this case.
Advanced Topics Guide
This is the table:
This select statement is used by transactions which perform the lookup:
WHERE COUNTY_CODE = :county_code;
This index was required to complete the primary key definition of the
The DBA realized that adding a new index which includes the
TAX_RATE column, in the low order of the key, would allow the lookup
to perform index only access, resulting in an I/O savings:
An alternative to composite indexes is creating separate indexes for each column. One
advantage is that these separate indexes may be used when processing the indexed
columns in a WHERE clause that specifies predicates that are OR’d together, since a
composite index cannot be used in this case. Another advantage is that if all the
columns are indexed separately, then you do not have to concern yourself with having
the leading columns of a composite index always being specified in WHERE
Suppose that the following set of queries access the STUDENT table, and
each query is contained within a number of transactions whose response time
is critical. These are the SQL query statements:
WHERE ID = :id;
WHERE NAME = :name
OR HOME_CITY = :home_city;
WHERE HOME_CITY = :home_city
AND MAJOR = :major;
WHERE MAJOR = :major;
Advanced Topics Guide
Refining the Physical Design
Since ID is the primary key of the STUDENT table, query number one will
never be a problem. A unique index must exist on ID in order for the primary
key definition to be complete.
A composite index on NAME and HOME_CITY can never be used by query
number 2, regardless of the column order, since the logical connector is OR.
This statement can only be satisfied by a table scan, or an index merge
between two indexes where the specified predicates occupy the high order
A composite index built on HOME_CITY followed by MAJOR could be
used by query number 3, but would not be acceptable for query number 4,
since MAJOR is not in the most significant position.
The following set of indexes could be considered for use by every one of the
above queries, and represents the most efficient way to satisfy the above
Good index candidates
The following list identifies some columns that may make good index candidates:
Primary key columns. By definition, primary key columns must have unique
indexes placed on them in order for the definition to be complete and allow
for the use of referential integrity.
Foreign key columns. These are usually very good index subjects for two
reasons. First, they are often used to perform joins with their parent tables.
Second, they can be used by the system for referential integrity enforcement
during delete operations against the parent table.
Columns that are referenced in WHERE predicates or frequently the subjects
of ORDER BY or GROUP BY clauses. Indexes on such columns can speed
up these operations significantly by preventing the creation of an
intermediate result table.
Any column that contains unique values.
Columns that are either searched by, or joined with, less than 5 to 10 percent
of all table row occurrences.
Advanced Topics Guide
Columns that are often the subject of aggregate functions (involving SUM,
AVG, MIN, and MAX).
Columns often used for validity editing in programs, where a value entered
by the operator is checked for its existence in some table.
Poor index candidates
Small tables
One common rule is to avoid creating indexes on tables which span less than five
pages. The costs associated with creating and maintaining these indexes are typically
greater than the cost of performing a scan of the entire table. Of course, a unique
index required to define a primary key in order to provide needed referential integrity
is an exception to this rule.
Large batch updates
Tables which experience updates in large batches usually have problems with index
overhead, particularly if the batches update more than 10% of the table. One possible
way to avoid these problems while still experiencing the benefits of indexes is to drop
all of the indexes on the table prior to the update, then recreate them after the update.
Skewness of keys
Another index pitfall is when the index’s key value distribution is significantly
skewed. In this situation, the cardinality of the index will appear to be quite high,
allowing the optimizer to compute a sufficiently low selectivity factor to make use of
the index often. In spite of this, the performance results of queries that utilize the
index are often unacceptable. This is because when there is a large percentage of rows
that have the same value for the index key, the index performance when accessing
these rows is much worse than when retrieving rows whose index key value occurs
less often. The following query will show the distribution of index values in a subject
SELECT key1, key2...keyn, COUNT(*) FROM tablename GROUP BY
key1,key2..keyn ORDER BY LastColumnNo
This query creates a report of index keys ordered from those occurring least often to
those occurring most frequently. If the difference between the occurrences at the top
of the list and those at the bottom is several orders of magnitude or greater, then this
index is probably going to cause performance problems when the rows being retrieved
fall into the more often occurring values. The fact that the total cardinality of this
index appears to be acceptable, and therefore its use is considered desirable by the
optimizer, makes this an insidious situation. It is only when the skewness of the keys
themselves is examined that the database designer discovers the performance
penalties that will accrue from the creation of this index.
Advanced Topics Guide
Refining the Physical Design
Other poor index candidates
The following column characteristics generally lead to indexes which are either too
expensive to maintain or of little benefit to system performance:
Low cardinality. These columns make poor index candidates since their high
selectivity factor influences the optimizer to avoid their use. Use of a low
cardinality index would often cause the DBMS to incur more I/Os than a table
scan, since many pages may be accessed multiple times.
Many unknown values (nulls). Null values represent a skew in the
distribution of what may otherwise be a column with acceptable cardinality.
Frequently changing values. Columns which are frequently updated cause
excessive index maintenance overhead. Each time the column is updated, the
DBMS must delete the old value from the index tree and insert the new value.
Also, columns which are updated by transactions that occur very frequently
may cause lock contention problems within the index structure. This is
particularly true if the column is a combination of date and/or time, or if the
column is a serially created key value. These situations cause transaction
activity to be concentrated in a narrow section of index values, which
increases the probability of lock contention occurring.
Excessive length. Single or composite index keys that are longer than 50
bytes will cause the index to grow to a large number of levels unless the
number of rows is low. Since at least one I/O is incurred for each level when
searching the index, this will degrade performance.
Database partitioning
The ability to partition a database across a number of database areas allows for the
throughput of the database to be greatly increased. Much of the benefit depends on
the way the operating system environment allows SQLBase to manage these database
areas. This varies depending on the platform on which the server is implemented
(such as NetWare, or OS/2). In general SQLBase will create a unique I/O thread for
each database area, which allows for a greater level of parallelism to be achieved in
the disk subsystem. Whether or not this will create bottlenecks at the disk controllers
or the drives themselves depends on the physical configuration of the I/O subsystem.
The other benefit realized from database partitioning is the ability to have databases
which are greater than the physical size of any individual disk drive. For very large
databases, there is no alternative to database partitioning since there may be no disk
drive available which is sufficient to store the entire database on a single drive.
The drawback to partitioned databases is that the physical disk location of a database
may be harder to identify. This becomes even more true if the DBA decides to allow
Advanced Topics Guide
Database partitioning
database areas to be shared by different storage groups, which are in turn used by
different databases. Monitoring free space in database areas is more difficult that
monitoring free space on a physical device, and determining the total free space
available to a storage group may involve the examination of a number of database
areas. Also, the DBA must be careful to maintain synchronization between the MAIN
database, which stores all of SQLBase’s database area extent information, and all of
the databases which are partitioned.
The bottom line to partitioning is that it is the best way to enhance throughput and
capacity in an appropriate operating environment, but it should only be undertaken
when the benefits are truly needed. Those DBAs who decide to implement database
partitioning must be prepared to spend time ensuring that their monitoring, backup,
and recovery procedures are all up to the task of handling the more complex details
associated with this environment.
When deciding how to partition a database consider the following factors:
Do not share database areas between databases and logs. Always specify
different storage groups for the data and for the log files, and ensure that no
database areas are shared between these two storage groups.
Place database areas intended for logs on a different physical device than any
of the database areas which are used for the data storage of tables. If a
database’s data files share a disk drive with their own log files, then the log
files may be damaged along with the database in a disk drive failure, and will
not exist to perform a full recovery of the database.
Consider not sharing storage groups between different databases. This
prevents databases from being interspersed in the same database area files, so
that if a failure occurs, the affected database will be more easily isolated.
If you are implementing database partitions in UNIX raw partitions, make
sure you have a technique developed to monitor space in the partition.
Do not set a default storage group unless you are absolutely sure that you are
prepared to manage all databases as partitioned.
The following query may be executed against the MAIN database using the
guest/guest account in order to ascertain available free extents in a database’s
database areas:
select a.name,b.stogroup,b.areaname,
@trim(c.pathname), c.areasize,
from syssql.databases a, syssql.stoareas b,
syssql.areas c, syssql.freeexts d
where b.areaname = c.name
and c.name = d.name(+)
and ((a.stogroup = b.stogroup)
Advanced Topics Guide
Refining the Physical Design
or (a.logstogroup = b.stogroup))
order by 1,2
Free space calculations
SQLBase provides a keyword, PCTFREE, to allow database designers the ability to
tune the amount of available space which is reserved within the data pages of tables
and indexes. Through the judicious use of this parameter, the designer may allow for
the growth of the database over a period of time without any degradation in
performance. Following this time period, a reorganization of the database should be
performed, marking the beginning of a new time interval for database growth. This
time interval, which is chosen by the designer, is termed the reorganization frequency.
The key considerations for choosing a reorganization frequency are the rate at which
the database grows and the availability of time in which to perform the
reorganization. If the database grows very slowly, then a great deal of time may be
allowed to pass between reorganizations without affecting its performance. On the
other hand, if the database is constantly growing, frequent reorganization may be
required to maintain consistent performance even with a generous amount of
freespace specified. Also, very large databases may take a great deal of time to
reorganize and this time may not be easy to find in the schedule of demand usage of
the computer system by the users. Finding the right trade-off between frequent
reorganizations and system availability can exercise the political ability of even the
most organizationally astute DBA.
The purpose of the PCTFREE parameter in the CREATE TABLE statement is to
reserve space in the database page for future growth of rows already allocated to that
page. No new row addition will be made to the page if it would cause the free space
on the page to drop below the PCTFREE value. You can ensure that some space will
remain on a page that will be available for the expansion of its rows.
This expansion can come about in three ways. First, a column of any data type may
be updated with data which is longer than what was previously stored. Second, a
column whose value was originally NULL is updated to possess a non-null value.
Finally, the DBA may decide to alter a table definition to add new columns to the
table which are then updated with non-null values.
Whenever any of these events occur and cause an existing row to expand on the page,
there must be sufficient free space on the page in order to avoid the creation of an
extent page. Whenever an extent page is created, additional I/O becomes required to
obtain rows which have spilled over onto that page. This additional I/O causes the
performance of the database to degrade. When enough of this activity has taken place,
with each data page having a string of extent pages connected to it, execution times
for transactions may be extended by several times their original objective.
Advanced Topics Guide
Free space calculations
When determining the appropriate values for the PCTFREE specification, consider
the following factors:
Tables which are read-only require no free space, and neither do indexes on
these tables. Specify PCTFREE 0.
If many table columns are often increased in length, specify a high (30-50)
PCTFREE value to accommodate this growth.
If a table is expected to have columns added to it after its initial creation,
specify a high PCTFREE to avoid the creation of many extent pages when the
new columns are added. However, if the new column will be populated with
non-null values only gradually, then specify a medium PCTFREE value (1030).
If transactions are experiencing problems with waiting for locks or deadlocks
due to a large number of active rows being located on the same page, the
concurrency for this table may be increased through the PCTFREE value.
Alter the table definition with a high PCTFREE value (80-90), then perform
a reorganization. The table will now have very few rows on each page
(possibly as few as one), which will decrease the likelihood of page locks
contending with the access of other rows. The downside is that the table will
occupy more room on the disk, and any table scan operations will take longer.
Following the implementation of the database, the DBA needs to observe its
performance over time to determine if any degradation is being realized prior to
reorganizations. If there is a significant performance drop due to excessive extent
pages being created, then appropriate action should be taken. The reorganizations
could be done on a more frequent basis. Alternatively, the PCTFREE specifications
could be increased prior to the next reorganization to stretch the effective performance
life of the database.
Advanced Topics Guide
Advanced Topics Guide
Chapter 7
Physical Design Control
This chapter describes a technique for performing physical database tuning on a
reiterative basis. The activities involved are:
Measuring or predicting transaction response times for a given database
design state.
Making a change to the design of the database in order to improve the
response time of one or more transactions.
Measuring or predicting the response time for all critical transactions with the
new design.
Deciding whether to accept or reject the database change based on the results
Advanced Topics Guide
Physical Design Control
As explained in chapter 1 of this section, database design is a six step process. These
six steps, which correspond to chapters two through seven of this section, are:
Build the logical data model, or the conceptual schema, of the ANSI/SPARC
three schema architecture.
Convert the logical model to a “first cut” physical database design which is the
internal schema in the ANSI/SPARC three schema architecture. This conversion
is accomplished by applying a few simple rules to the standard Entity
Relationship diagram.
Define the external schema, or DBMS views.
Define the database transactions.
Refine the “first cut” physical database design by applying some heuristics to
common situations.
Perform physical design control.
Physical design control is the iterative process whereby the physical database design
is adjusted, or tuned, so that the execution time of a database transaction, which is not
meeting its performance objective, is decreased. The goal of physical database design
is to ensure that every database transaction meets or exceeds its performance
objectives. (In reality, it is sometimes prohibitively costly for every transaction to
meet or exceed its performance objectives. In these cases, the goal of physical design
is to ensure that every high priority transaction meets its performance requirements,
and all other transactions do not deviate unacceptably from their performance
requirements.) Consequently, physical design control is complete when every
transaction meets or exceeds its performance objectives.
This chapter presents the procedure of physical design control. In addition, common
tuning scenarios are discussed along with tips and techniques which may be used to
speed up specific types of transactions.
The process
The process of physical design control is accomplished in three steps:
Determine the execution time of each transaction and locate the highest priority
transaction which is not meeting its performance objectives.
Adjust the physical constructs of the database so that the transaction takes less
time to execute.
Advanced Topics Guide
The process
Go back to step one, unless all transactions are now performing within their
required response times.
Each of these steps are discussed in the following sections.
Step one
The execution time of individual transactions may be determined by actually
executing the transactions or by estimating the execution time using some forecasting
model. The former method is referred to as the “brute force method” and the latter as
the “forecast method.”
Brute force method
To apply the brute force method, the database designer must compile the physical
schema and load the database tables with data. Once this is complete, SQLTalk scripts
or application programs may be developed which submit each transaction to
SQLBase and record the execution time.
However, in order for the results to be meaningful, the database size (in terms of the
number of rows in each table) must be approximately equivalent to the size of the
production database. This is because the query optimizer chooses the access path
based on statistics kept in the catalog. If the database is relatively small, as compared
to the production database, it is likely the access path chosen by the optimizer will
differ from the one which would have been chosen had the transaction been run
against the production database.
Therefore, for the results of the brute force method to be meaningful, data volumes
which approach the size of the production database must be loaded in the database.
This fact limits the usefulness of the brute force method to relatively small databases.
(Of course, “small” is a relative term. On a 386/33MHz computer running DOS,
small probably means less than 50MB. On a 486/66MHZ computer running a 32 bit
operating system like NetWare, a small database is probably less than 150MB.) This
reasoning centers on the time required to process large volumes of data.
For example, consider a 500 megabyte SQLBase NLM database running on a 486/
66MHz Compaq SystemPro. It takes several days of processing time just to load a
500 megabyte database on this hardware platform. Furthermore, a database of this
size typically contains at least one table with several hundred thousand rows. On a
486/66MHz computer, it requires several hours of processing time to build one index
on a table containing several hundred thousand rows. It is not uncommon for a 500
megabyte database to have dozens of tables, each containing several indexes.
Perhaps if the database tables and indexes only have to be loaded or built one time, the
logistics and computing time necessary to handle a large database would not be
insurmountable, but this is not the case. As we shall see in step two, physical design
control involves altering the physical constructs of the database to decrease execution
Advanced Topics Guide
Physical Design Control
time. For example, several different indexes may have to be tried before a specific
transaction meets its performance objectives. Similarly, the database designer may
choose to cluster a table on a different key to see if performance increases. To
accomplish this, the table must be dropped, the data reloaded and all the indexes
rebuilt. Sometimes a design iteration may require the data load programs to be
altered. This can occur, for example, if the database was denormalized in an attempt
to increase the performance of a transaction. Of course, re-coding application
programs may involve an enormous amount of effort.
Consequently, each design iteration in the physical design control process potentially
requires the data load procedures to be repeated. On a complex database, ten or more
design iterations may be required before all the transactions are meeting or exceeding
their performance objectives. Hence, the database tables and indexes may have to be
built ten or more times. Therefore, as stated above, the usefulness of the brute force
method is limited to relatively small databases.
Forecast method
The forecast method is ideally suited for medium to large databases since many
different design strategies may be tried without having to actually load data into the
database. This technique involves estimating the execution time of a particular
transaction through the use of some forecast model.
Forecast models typically try to predict how many physical block reads or writes and
CPU cycles a specific transaction will consume. These numbers are then multiplied
by the average execution time of a physical I/O or CPU cycle. For example, most disk
drives commonly used today in Intel-based database servers have a total data access
time of less than 20 milliseconds. This is the average time required to position the
read/write head on the desired track plus the average rotational delay time (the time
required for the desired sector to spin around to the read/write head). According to
this model, if a given transaction required 1000 physical I/O, on average the
transaction will execute in twenty seconds (1000 * .02). We are ignoring CPU time
here for the sake of simplicity. In fact, CPU time is often omitted from forecast
models because CPU time is usually a small percentage of overall transaction
execution time.
Of course, these forecast models are an over-simplification of true performance and
overlook many factors such as disk controller caches, operation system file buffers,
and database page buffers. However, the measurements are usually sufficiently
accurate for the desired purpose. It is not necessary to know exactly the execution
time of a given transaction. If the forecast is only 50% to 75% accurate, we still have
an adequate measuring rod to perform physical design control. The estimate of
execution time provides a relative measurement, as long as the estimating algorithm
is applied evenly. This relative measure can then be used to identify those transactions
which are expected to grossly exceed their performance objectives.
Advanced Topics Guide
The process
Before the relational model and SQL became the de facto standard, forecast models
were used almost universally by database designers. This is because most DBMS of
that time employed a navigational database manipulation language (DML). With
these navigational DMLs, the programmer determined the access path for each
database access. Once the access path is known, it is relatively straightforward to
estimate physical I/O. Hence, forecast models were relatively easy to develop.
However, with the rise of the relational DBMS and SQL, forecast models have
become much more difficult to build. This is because SQL is a declarative DML. With
SQL, the programmer does not determine the access path. This task is delegated to the
query optimizer. So, in order to accurately forecast the number of physical I/O, the
forecaster must first “guess” which access path the query optimizer will employ for a
given database operation. The cost model of query optimizers are extremely complex
and beyond the reach of most database designers.
The Centura advantage
Fortunately, SQLBase provides an elegant solution to the problem of forecasting
execution time. SQLBase contains a number of features which, when used together,
allow the SQLBase query optimizer to serve as the forecast model.
First, the Explain Plan facility has been enhanced to provide:
More detailed information describing the access path
The “cost” of the access path
Consequently, when in “explain mode,” SQL commands are not actually executed or
result rows returned. Instead, the query optimizer determines the lowest cost path and
returns a description of that path plus the cost. The cost returned by the query
optimizer is the optimizer’s own projection of the execution time of the SQL
command. The cost is calculated by estimating the number of physical I/O and CPU
cycles required to actually execute the SQL command based on various database
statistics. Please refer to section three of this manual for a detailed discussion of the
Since the query optimizer uses current database statistics, care must be taken by the
database designer to ensure that these statistics reflect the production environment.
Otherwise, the query optimizer might choose a different path or return an
inappropriate cost.
One way to ensure that the query optimizer is using the right statistics is to load the
database with production data volumes. However, this approach suffers from the same
limitations as the brute force method. Fortunately, SQLBase also provides several
new commands which allows the database designer to override the statistics with user
calculated values.
Advanced Topics Guide
Physical Design Control
In SQLBase, all statistics used by the query optimizer are now stored in the catalog as
columns of the various system tables automatically created and maintained by
SQLBase. In addition, the UPDATE STATISTICS command may be used to store
user-desired statistics in these columns in order to influence the choices made by the
Consequently, the procedure which the database designer should follow when
attempting to use the query optimizer as a forecasting tool is:
Compile the physical schema DDL.
For each table and index, estimate the values for the optimizer statistics
which will be stored in the catalog once production data volumes are loaded.
Store these user-determined statistics in the catalog using the UPDATE
Put the database in explain mode using either the sqlexp function or the
Submit SQL commands to SQLBase in the usual way and record the access
path and cost returned by the query optimizer.
Calculate the total forecasted cost of the transaction by summing the cost of
each SQL command in the transaction.
Step two
During step one of physical design control, the execution time of each transaction is
determined and the highest priority (during transaction definition, each transaction is
ranked in terms of its relative priority and its performance requirements specified)
transaction which does not meet its performance objectives is located. Thus, the input
to step two is one transaction. The objective of step two is to adjust the physical
constructs of the database so that this transaction’s execution time is decreased. This
is often referred to as “tuning the schema” for a transaction.
Most database transactions consist of many SQL commands. Consequently, the first
task in step two is to determine which of the SQL commands that make up the
transaction are contributing the most to the overall cost of the transaction. The
designer should then focus his attention on these commands first.
Once a specific SQL command is targeted, the database designer may then begin to
consider alterations to the physical schema or the SQL syntax which may increase
performance. The general design goal is to reduce the number of physical I/Os
required to satisfy the SQL command. Three fundamentally different alterations may
be made to the physical schema:
Add or change row clustering, primarily via hashing
Add or change an index
Advanced Topics Guide
The process
Denormalize a table
Similarly, the SQL syntax may be changed in two fundamental ways:
Restrict the selection criteria
Replace the SQL command with an equivalent one
Add or change row clustering
The mechanics of the clustered hashed index are covered in chapter 10 of this manual.
You are encouraged to review this section. In addition, general principles for
determining whether a hashed table will result in improved performance in a given
situation was discussed in the previous chapter.
Also, note that in some circumstances, transaction performance may be increased by
unhashing the table and clustering it through a different technique. For example, one
could improve performance during table scans for tables which contain static data
(data rows which are never updated or deleted) by unloading the data, physically
sorting the rows in some specific order, and inserting the rows in this specific physical
Add or change an index
Determining the optimal index set for a given database and transaction is perhaps the
most difficult task the database designer faces and it requires a thorough
understanding of the optimizer. In addition, metrics are presented in the previous
chapter to evaluate the quality of a proposed index as well as guidelines to follow
when creating an index.
Denormalize a table
Denormalizing a database is a common technique to improve the performance of a
SQL command. Denormalization refers to the process of converting a database which
is in third normal form to some lower level of normalization for performance reasons.
The concepts of first, second, and third normal form are presented in chapter two and
you are encouraged to review this material.
While in the majority of situations, the benefits of direct implementation of
normalized data structures outweigh any disadvantages, in a few cases there may be
sufficient justification to perform denormalization in a controlled, deliberate manner.
For example, whenever many joins will be required across large tables on a frequent
basis, the performance penalty of these join operations may be lessened by combining
some of the tables or introducing selective data redundancy among the tables. This
denormalization must be done with careful consideration to the costs involved in
working around update and deletion anomalies, performing processing to update
redundant data, and increasing the recurring maintenance costs for the database. Of
course, the benefits realized from the denormalization include fewer join operations
Advanced Topics Guide
Physical Design Control
and fewer foreign keys along with their usual indexes. When these benefits are more
critical to the success of the system than the drawbacks of a non-normal database,
then denormalization is done as a part of the physical database design process.
Denormalizing with prime attributes is one common form of denormalization.
This involves redundantly storing a “prime” (meaning not derived or calculated) data
element in both a primary table and a secondary table. The primary table is the table
where the data element “belongs” in terms of the normalization rules. By storing the
data element on the secondary table, the normalization rules are violated.
For example, consider the query, “display a list of all customer zip codes for which
there currently exist open leads, and rank the zip codes in descending order by the
number of open leads.” The SQL command for this query might be:
Note that the ZIP_CODE of the customer is stored in the CUSTOMER table because
that is where the attribute belongs. For any one CUST_ID, there exists only one
ZIP_CODE. (We are ignoring here the fact that a company may have many locations
and each location may have a postal and delivery zip code.) However, to execute this
query, the LEAD table must be joined with the CUSTOMER table (see the WHERE
clause above). If the CUSTOMER table contains several hundred thousand rows and
at any one point in time there exists only a relatively small number of open leads, then
the performance of this query could be significantly increased by denormalizing the
LEAD table.
To denormalize the LEAD table to speed up this query, the customer ZIP_CODE
field is redundantly stored in the LEAD table. The new SQL command to produce the
same result is:
Thus, the join of the CUSTOMER table with the LEAD table is avoided resulting in
increased performance of the target query. However, the LEAD table is now not in
third normal form because the ZIP_CODE attribute is not dependent on the entire
primary key of the LEAD table. Presumably, the primary key of the LEAD table is
(CUST_ID, LEAD_NBR). Given this, the ZIP_CODE attribute depends only on the
CUST_ID; regardless of the LEAD_NBR, the customer’s ZIP_CODE remains the
same. Thus, the LEAD table now violates the second normal form rule.
Advanced Topics Guide
The process
Violating the second normal form rule, in this situation, does increase performance of
the target query. However, this increased performance comes at the price of an update
anomaly. We cannot now change the ZIP_CODE of a customer in the CUSTOMER
table without changing every lead for that customer. Hence, the update customer
transaction has to be written so that any change of a customer’s ZIP_CODE is
reflected in all leads (otherwise the data in the LEAD and CUSTOMER tables would
be inconsistent). Denormalizing the LEAD table to speed up the target transaction
also resulted in:
Decreased performance of the update customer transaction.
Placing the data integrity of the CUSTOMER and LEAD tables in the hands
of the application.
This particular example of denormalization results in an update anomaly. Insertion
and deletion anomalies are also possible. (See C.J. Date, An Introduction to
Database Systems, Addison-Wesley for a detailed explanation of update, insertion
and deletion anomalies.) Every time a table is denormalized, the database designer
must be aware of, and devise a strategy to deal with, update, insertion, and deletion
anomalies. In addition, one must rely just a little bit more on the application programs
to enforce the data integrity of the database. If the programs contain bugs relating to
their handling of these anomalies, data integrity will be compromised. For these
reasons, denormalization should only be considered as a last resort. The database
designer should try all other alternatives to speed up the query before resorting to
Further, one should resort to denormalization only in the case when a specific
transaction cannot otherwise meet its performance objectives. Some database design
methodologies suggest that the designer denormalize the database as a general rule. In
other words, certain tables are denormalized based on broad guidelines before any
transactions are timed. The table is thus denormalized whether it is needed or not. The
only valid reason to denormalize a table is to achieve increased performance of a
given transaction that cannot otherwise be obtained. To denormalize a table before it
has been demonstrated that the transaction performance objectives cannot otherwise
be met is reckless given the cost of dealing with update, insertion, and deletion
anomalies or a loss in data integrity.
Denormalizing with aggregate data is another common form of
denormalization. This deals with adding aggregate or derived data attributes to
database tables. Aggregate data are data items derived from other lower level data and
are frequently required in decision support and executive information systems in
order to satisfy the large degree of summary and exception reporting these type
systems provide.
For example, consider the query “produce a list of salesmen and their total year-todate sales in descending order from the top selling salesman to the salesman with the
Advanced Topics Guide
Physical Design Control
least sales.” To produce this report, all orders for the current year must be read. Most
order entry systems store order data in two tables: the ORDER_HEADER and
ORDER_DETAIL. The ORDER_HEADER table contains data germane to the entire
ORDER_DETAIL table stores information pertaining to one order line item.
Consequently, to produce this report, the ORDER_HEADER table must be joined
with the ORDER_DETAIL table. The SQL statement to produce this report is:
If there are 50,000 rows in the ORDER_HEADER table and 300,000 rows in the
ORDER_DETAIL table (on average each order has six lines items), the performance
of this query can be significantly improved by adding an aggregate data element,
TOTAL_ORDER_AMOUNT, to the ORDER_HEADER table. Whenever a row in
the ORDER_DETAIL table is inserted, updated or deleted, the
TOTAL_ORDER_AMOUNT is decremented or incremented appropriately so that
this summary field contains the total sales amount of the entire order (which is the
sum of all the line items). The SQL command can then be written to avoid the join:
Technically, the ORDER_HEADER table is now in violation of the normalization
rules and update, insert, and delete anomalies exist. The create/modify order
transaction must be written so that this total field in the ORDER_HEADER table is
maintained as order lines are added, changed, or deleted.
When denormalizing by adding aggregate data to the database, the decision the
designer must make is whether to allow SQLBase or programs to calculate these
values each time they are needed, or to actually store them in the database so that they
can be queried without the overhead incurred from their derivation. The key factors
involved in this decision are the frequency of access for these data items, the response
time required, and the frequency and types of update operations that are performed on
the lower level data from which the aggregate data is derived.
Consider the case of the student’s GPA. This information is probably needed for each
academic inquiry transaction that the system performs, and these transactions may be
executed quite frequently and require a reasonably low response time (probably two
seconds or so). The lower level data, which are the grades a student has received in
each course he has taken, could be quite large for a student with a long academic
career. Not only would the on-line inquiry program have to retrieve all this data to
perform the calculation, but the logic needed is also quite complex.
Advanced Topics Guide
The process
Furthermore, this grade data is updated infrequently, probably only at the end of the
semester. This update could be performed with a mass insert program that can be
designed to perform GPA calculations for students in an efficient manner. On-line
transactions which perform grade changes are not highly used and can suffer from
some slower response time while the program is calculating and updating the GPA.
This is an excellent scenario for allowing aggregate data to be placed in the design,
since the benefits realized will outweigh the costs.
The example of a customer’s total sales for a year probably falls on the other side of
the trade-off. This data is usually not used for on-line inquiry transactions, but more
often appears on reports for sales or marketing departments. As such, there is little
benefit from speeding up the derivation of this data. Also, it is easily computable by a
SQL query which performs a GROUP BY operation. The values from which the
information is derived are the sum of all customer order total sales, which could be
frequently updated for active customers. Every order transaction that is processed
through the system would have to perform the updating of the customer’s total sales,
which would cause overhead in a transaction which probably has stringent response
time and throughput requirements. When everything is considered, this data item
should not be placed in the customer entity and those programs needing it will have to
pay the performance penalty of performing the aggregation query needed to derive its
Each case of aggregate data should be subject to the same level of analysis as
demonstrated previously. Only in those cases where there are clearly significant
benefits from actually storing the derived data should it be maintained in the database.
The general rule of thumb should be to eliminate derived data and, as with the
denormalization trade-off, whenever the rule is broken, the results of the analysis that
lead to that decision should be thoroughly documented within the database design.
Thus, this form of normalization has the same pitfalls as redundantly storing prime
attributes. However, unlike prime attributes, derived attributes add information to the
database. Consequently, denormalizing with aggregate data is perhaps preferable to
redundantly storing prime attributes.
Restrict the selection criteria
In addition to changing some physical construct of the physical schema, the database
designer may also change the SQL command in order to increase the performance of
a transaction. The first type of change the database designer can make to the SQL
command is to further restrict the selection criteria. In order to do this, the database
designer takes advantage of some knowledge of the application in order to add an
additional restriction to the selection criteria.
For example, consider the query, “produce a report of all open orders.” Now also
suppose that:
Advanced Topics Guide
Physical Design Control
Orders are either “open” or “filled.”
On average, 300,000 orders exist (four years history plus the current year).
The SQL command to produce this report is (we ignore the ORDER_HEADER,
ORDER_DETAIL data structure as it is not required to illustrate the point):
The optimizer is forced to execute a table scan on a table containing 300,000 rows to
satisfy this query. The only index which could be used for this query would be an
index on the ORDER_STATUS attribute. However, this attribute has a cardinality of
two (meaning it has two distinct values); hence, an index on this attribute would have
a selectivity factor so high the optimizer would never choose it.
Now suppose that the database designer also knows that:
On average 240 new orders are added every day.
99.9% of all orders are usually filled by 12:00pm the same business day on
which they are received or the following day.
99.9% accuracy is sufficient for purposes of the report.
The problem is that the optimizer must read 300,000 rows to find the 240 whose
status is “open” (the ones not filled). Since 99.9% of all orders are received and filled
on the same business day and the user’s requirements can be met with this
approximation, an index could be placed on the ORDER_DATE attribute and the
query could be rewritten as:
Since four years of order history is maintained in the ORDER table, the cardinality of
the ORDER_DATE would be high and the selectivity factor of the index extremely
low. Hence, the optimizer would probably use this new index to read the 240 open
orders directly.
The point of this example is that the database designer took advantage of some
knowledge of the application to apply a further restriction to the selection criteria
which, because of the nature of the system, produced the identical results. The
implication of this observation is that the database designer must understand the
application or these opportunities will be missed.
Advanced Topics Guide
The process
Replace the SQL command with an equivalent one
SQL is a powerful and flexible database manipulation language; the same result set
may be produced with several different SQL commands. A subquery may be
expressed as a join. Several OR predicates may be replaced with one IN list. For
example, the following SQL statement:
OR COL1 = ‘B’
OR COL1 = ‘C’
may be rewritten as:
WHERE COL1 IN (‘A’,’B’,’C’)
Equivalent SQL commands sometimes produce significantly different performance
results. A SQL command may run much faster than one of its equivalent forms. The
reason for the difference in performance is that the SQLBase query optimizer is able
to detect certain conditions for which it is able to apply some fast access path
techniques. With other forms of equivalent SQL commands, the optimizer is not able
to detect these conditions.
The details of these circumstances require a detailed understanding of the query
optimizer. You are encouraged to review section two of this manual which discusses
the SQLBase query optimizer at length. However, in practice, when a given SQL
command is executing too slowly, the database designer should re-code the command
in as many different equivalent forms as possible to see if one form produces a
significantly different performance result. The implication of this is that the better and
more experienced the database designer is with SQL, the more equivalent forms of a
given SQL command the database designer will be able to try.
Advanced Topics Guide
Advanced Topics Guide
Chapter 8
Database Pages
In this chapter we:
Introduce the concept of the database page, which is the basic unit of disk
storage used by SQLBase.
Cover the following database page types in detail:
Data row pages
Data extent pages
Long varchar pages
Overflow pages
Index pages
Control pages
Discuss SQLBase’s page allocation and garbage collection algorithms.
Advanced Topics Guide
Database Pages
Pages and page types
The fundamental data structure of a SQLBase database is the database page. The
SQLBase database file (the .dbs file) is composed of many pages (see Figure A). The
SQLBase page size is fixed; all pages have the same length. The Windows, OS/2,
Windows NT, and NetWare versions of SQLBase use a page size of 1024 bytes (or
1K). In most DBMS, database pages are a multiple of 1K to optimize reading and
writing to secondary storage devices. Disk controllers read and write data in blocks
that are usually multiples of 1K. In addition, most operating systems’ file buffers are
multiples of 1K. For example, the default block and cache buffer size for NetWare is
4K. If the database page size is equal to, or a multiplier of, this common 1K block
size, I/O is optimized since one physical I/O can read or write one or more database
P hy sic al
P age N o
P hy s ic al
P ag e N o
Phy s ica l
Pa ge N o
1 024 bytes
1024 bytes
tota l
10 24 bytes
Logical File Offset
P hys ica l
P age N o
1024 bytes
( N-1)1024
Figure 8A - The SQLBase database is a set of 1024-byte pages.
The relationship between the size of the database page and block and file buffer is an
important consideration in access path selection (which is discussed at length in
chapters 15 through 18). For example, assuming the default block and file buffer size
in NetWare, one physical I/O actually retrieves four database pages. Thus, SQLBase
can retrieve four database pages with one physical I/O for sequential reads (such as
during table scans). This assumes the next physical page is also the next logical page
which for sequential reads is reasonable. On the other hand, random reads require one
physical I/O for each logical read request (to retrieve a row). Since one database page
can store more than one row, a table scan is a much more efficient access method if
more than just a small percentage of rows is needed.
Each database page is assigned a physical page number which is four bytes long. This
number is assigned sequentially: the first page in the database is assigned page
number one, the second page number two and so on. SQLBase uses the physical page
number, along with the page size, to calculate the logical file address of each page.
For example, the page with the physical page number of 10 would begin at byte
number 9,216 (the formula is (n-1)*1024) within the database file and continue for a
length of 1024 bytes. This is referred to as the logical file offset. SQLBase requests
Advanced Topics Guide
Pages and page types
pages from the operating system by specifying logical file offsets. With the exception
of Unix raw devices, it is the job of the file system and disk controller to convert this
number into a physical storage device and the actual track and sector (block) where
the database page is stored, using the specific device geometry information coded into
the driver module supplied by the manufacturer of the disk subsystem.
Database pages are also linked to form lists of pages. These lists are referred to as
linked lists. Each page contains a pointer to the next, and sometimes prior, page in the
list (see Figure B). The pointer consists of the physical page number and facilitates
the sequential access of all the pages within a specific list.
Ph ys ica l
Pa ge N o
P rio r N ext
P ag e P a ge
P hy s ica l
P age N o
P rio r N ex t
P ag e Pag e
P hys ic al
P age N o
0 15
1 27
P rio r N ex t
P a ge P ag e
Figure 8B - Database pages are linked together in lists known as linked lists.
Each database page is a member of one page type. The basic page types are data
pages, index pages and control pages. Data pages contain actual table row data.
SQLBase has four types of data pages: row pages, extent page, long varchar pages
and overflow pages. Index pages are used to store B+Tree indexes. Control pages
maintain information for internal purposes such as which pages are empty or full.
Each of these page types is discussed in detail in the following sections.
A database page contains data only for one page type. For example, BTree pages only
contain data (consisting of symbolic keys) for a particular BTree index. Likewise, row
pages only contain data.
A database page also belongs to one and only one group and contains data for only
one group. A group is a database object such as a table or index, and a collection of
pages of different page types. For example, a table is made up of n number of row
pages, extent pages, LONG VARCHAR pages, and control pages.
Advanced Topics Guide
Database Pages
Each database page is also assigned a logical page number. The logical page number
is assigned sequentially within the group: the first page within the group has the page
number of 1, the second page in the group 2, and so on.
Ty pe
Gr oup Log. Prio r N ex t
Ptr Pag e Pa ge
28 1
0 15
1 st p ag e in the custom er table
T y pe
Group Log. Pri o r Ne x t
Ptr Pag e Pag e
15 Data 28
1 27
2 nd pa ge i n the cu stom e r table
Typ e
Grou p Log.
P tr
Pri or N ex t
Pa ge Page
15 0
3 rd page i n t he c ust om er t able
Figure 8C - All the pages of a specific group form a separate linked list. Each group
has one page type and stores data only for that page type.
Data pages
As mentioned previously, SQLBase contains four basic types of data pages: row
pages, extent pages, LONGVARCHAR pages, and overflow pages. Each of these
pages types is discussed in the following sections.
Row pages
When a table row is originally stored in the database, it is stored on a row page. The
specific location occupied by a database row on the page is called a slot. The number
of slots that may reside on a page is a function of the size of the rows and the page
size. SQLBase keeps track of the location and size of each slot within a database page
through the use of a slot table. The slot table serves as an index to all the slots
contained within a database page.
Each row page contains a generic page header, row page header, slot table, variable
number of slots and a page footer. The slot table grows down the page from the row
page header and the slots grow up from the page footer. The freespace within the page
is between the slot table and the last slot added (see Figure D).
Page header
The page header is 53 bytes long and contains the physical page number, page type,
group number, logical page number, next and previous page pointers, and a log record
pointer (see Figure E). The log record pointer points to the record in the log that
contains the before and after image of the page for the last SQL operation that
modified the page.
Advanced Topics Guide
Data pages
Figure 8D - Row pages contain a page header, row page header, slot table, variable
number of slots and a page footer. The free space within a row page is between the
last entry in the slot table and the last slot.
Row page header
The row page header is 16 bytes and contains the slot size (the number of entries in
the slot table), available freespace (in bytes) and a pointer to the next (or first) extent
Slot table and slots
Every table row is stored in a separate slot. Each slot contains a row header followed
by the set {column identifier, column length, value} for each non-null (and nonLONG VARCHAR) column in the row. The row header contains the row serial
number, update serial number, and the number of columns in the row. Each slot has a
corresponding entry in the slot table. A slot table entry is a two byte field. The first bit
is used as an empty or full flag. The bit is on if the slot contains data; the bit is off is
Advanced Topics Guide
Database Pages
the slot is empty. The remaining bits contain the byte offset of the slot within the page
(see Figure E).
P ag e H e ad e r
P h ys ic a l P a ge G rou p L o g ic a l
P re vio u s
P ag e N o. T yp e N u m b er P ag e N o. P a g e N o.
R ow P a g e H e a de r
N e xt
Lo g
P a g e No . P o in ter
S l ot
S iz e
F re e
S pa c e
E xten t
Page No.
S lo t T a b le
E n try 1 En try 2 E n try 3
9 01
8 61
Fr eespace
S l ot 3 (40 b ytes )
86 1
9 00
S lo t 1 (1 0 b yte s )
P a ge F oo te r
L o g Po in ter
10 20 1 02 1
1 02 4
S lo t 2 (2 0 b yte s )
R ow He a d er
R o w S eria l N b r
U pd a te S e rial N b r
N b r C olu m ns
1 10Fre d Sm i th 2 7A tl an ta
3 2G A
D a ta
L e ng th
C ol ID
92 0
90 1
Figure 8E - Each table row is stored in a slot. The slot contains the slot header and
the set {column id, length, data value} for each non-null (and non-long varchar)
The row serial number is a unique identifier assigned to each row in a table. The row
serial number is assigned sequentially: the first row inserted into the table is assigned
the row serial number of one, the second row is assigned two, and so on. The update
serial number is a unique identifier assigned to every row in a common UPDATE or
INSERT command. The update serial number is also assigned sequentially: all the
rows updated/inserted with the first UPDATE/INSERT command are assigned an
update serial number of one, all the rows updated/inserted with the second UPDATE/
INSERT command two, and so on. The update serial number is used to prevent lost
updates with the Cursor Stability and Release Lock isolation levels (see chapter 13)
and infinite looping with UPDATE and INSERT commands.
Consider the SQL UPDATE command:
With this command, if the DBMS did not flag the rows in table y that met the where
clause condition before the update command began, then the update command will
Advanced Topics Guide
Data pages
result in an infinite loop. The update serial number is used for this purpose in
SQLBase, which contrasts with some other RDBMSs which disallow the placement
of columns within the WHERE clause of an UPDATE statement which affects their
Page footer
The page footer redundantly stores the log page pointer. This field serves as a
consistency check and is used during transaction recovery to protect against partial
page writes. Many Intel-based computers traditionally block hard disks with 512 byte
sectors. This means the hard disk controller reads and writes in units of 512 bytes.
However, the SQLBase page size is 1024 bytes. Thus, it is theoretically possible on
such a computer for only a portion of a database page to be written to disk (when an
undetected disk error occurs after the first sector is written but before the second 512
bytes sector is written). Thus, by storing the log page pointer in both the page header
and footer, SQLBase can detect when this situation occurs and respond accordingly.
SQLBase uses a rowid for all external references to table rows. A rowid is composed
of the physical page number, slot number, row serial number and the update serial
number. The physical page number is used to locate the row page containing the row.
The slot number is used to locate the specific entry in the slot table that contains the
offset of the slot (see Figure E).
The slot table/slot data structure is an indirection technique that allows a specific slot
to move within a page without affecting external references (namely rowids). It is
necessary to relocate slots within a page after the deletion of rows to reclaim
freespace and avoid fragmentation of the page (see page garbage collection later in
this chapter).
The row serial number is required to verify the row is actually stored in the specified
slot. Since slots are reused, the row could have been deleted and another row stored in
the same slot. The row serial number allows SQLBase to determine when a row
pointed to by a rowid has been deleted. Likewise, the update serial number allows
SQLBase to determine when a row referenced by a rowid has been updated by
another user.
Extent pages
Extent pages are extensions to row pages and are used to store some of a table row’s
columns when all of the columns will not fit in one row page. Columns can become
displaced in two circumstances, the most common of which occurs when column
values are updated such that the length of the columns increase significantly in size.
As mentioned previously, when a table row is originally inserted into a table, all
columns (except LONG VARCHAR columns) are stored in a row page (assuming the
Advanced Topics Guide
Database Pages
row will fit into one page). However, later the row may be updated resulting in an
increased row length. This can occur, for example, when a column that was null at the
time of insertion is updated with a non-null value. When the row is updated such that
its length is increased, sufficient freespace may not exist on the row page (the current
base page) to store the new column values. When this situation occurs, SQLBase
stores the new column values on an extent page.
The second situation which can cause columns to be displaced from their base page
occurs when you attempt to insert a row with a row size greater than the usable row
page size. Usable page size is the page size minus the sum of the page size reserved
for later row expansions (through the use of the PCTFREE specification on the
CREATE TABLE statement) and the page overhead. For example, for the NLM
version of SQLBase, the page size is 1024 bytes and the page overhead is 86 bytes.
Consequently, some columns will be displaced from the base page and stored on
extent pages when you attempt to store a row with a row width of greater than 938
bytes, assuming the PCTFREE parameter is zero.
Extent pages, like row pages, contain a page header, row page header, slot table,
variable number of slots and a page footer. However, the next and previous page
pointers in the page header are not used; extent pages are linked together and to the
base page in a separate linked list using the extent page pointer in the row page header
(see Figure F).
Base R ow Pages
T ype
G roup
Exte nt
P a ge
Gr oup
Ex tent
Pa ge
Extent Pages
S lot 2
E xtent
Gr oup
Ex tent
S lot 2
S lot 2
S lot 2
Figure 8F - Every row page may have up to ten extent pages which are linked to the
base row page. Extent pages share the same group number with the base page.
Displaced columns are stored in the same slot as the columns on the base page.
Advanced Topics Guide
Data pages
When an update or insert occurs such that the entire row cannot be stored on a base
row page, SQLBase searches for the first extent page with sufficient freespace to
contain the entire displaced columns. If no extent page is found which meets this
criteria, SQLBase allocates one new extent page (up to a maximum of ten).
A given column, consisting of the set {column identifier, length, value}, is never split
across pages. The slot in the base page contains no reference to the extent page or the
displaced columns. The displaced columns are stored in the same position in the slot
table in the extent page as the row to which they belong is stored in the base page.
This means that a row’s slot number remains unchanged from the base page value
regardless of how many extent pages it may span. When retrieving a row, SQLBase
first retrieves the base page and accesses all columns stored in the slot. If some
columns are displaced, SQLBase searches sequentially through the chain of extent
pages until all displaced columns are found.
Out of extent pages error
The 00802 ROW ISA “Insufficient (data page) Space Available” error occurs when a
given row cannot fit within a row page and ten extent pages have already been
allocated to the row page but none contains sufficient freespace to store the displaced
columns. The 00802 error occurs most frequently during update commands where
many columns in the row were originally inserted with null values. If this situation
occurs infrequently and in conjunction with specific tables, the appropriate remedy is
simply to reorganize the placement of the table rows on the database pages by
exporting and importing the table. However, if the situation occurs frequently, this
indicates that the nature of the data is such that rows are growing wider at a significant
rate. In this case, the appropriate remedy is to adjust the PCTFREE parameter on the
CREATE TABLE statement. The percent free parameter will cause SQLBase to
reserve more freespace for expanding rows.
The 00802 error can also occur during the insertion of rows. If this situation occurs
frequently, the average row width is simply too large and the database should be
redesigned. One could define a table with 255 columns and attempt to store 254
characters in each column for a total row width of 65,025 bytes. This insert command
would exceed the 10,318 total bytes available in the base pages plus ten extent pages
(computed by multiplying the usable page space, 938, by 11) and result in the 00802
error. Because in most environments SQLBase uses 1K pages (the exception is in the
UNIX implementations of SQLBase, which use a 2K page size), table width should
be kept to less than 938 bytes (usable page space after overhead) to optimize row
retrieval time.
Effect on performance
Extent pages negatively affect performance by requiring additional page I/O to access
row data. If a row is stored without the benefit of extent pages (fitting completely on
one row page), then the row can be accessed with one physical I/O. However, if the
Advanced Topics Guide
Database Pages
row is spread over one row page and one or more extent pages then multiple I/Os may
be required. Since extent pages are allocated on an as needed basis, it is unlikely the
extent pages for a given row page would be stored in the same physical block (except
following a reorganization). However, it is possible several extent pages all belonging
to the same row page are stored in the same physical block.
For example, assume a “narrow” row (meaning one with a row width less than the
usable row size minus reserved space) is inserted into the database and stored either
on an existing page with sufficient freespace or a new page which has no extent pages
allocated at the time of the insert. Now assume that at some later point in time the row
is updated with increased column widths which must be stored on an extent page and
that SQLBase allocates a new extent page for this purpose. At this time, it is very
likely that the pages physically adjacent to the base page have already been allocated
for other purposes (such as additional row pages or extent pages for other base
pages). Consequently, the page which SQLBase does allocate for the extent page will
not be physically adjacent to its base page.
On the other hand, consider the following. Assume a “wide” row (meaning one with a
row width greater than the usable row size minus reserved space) is inserted into the
database. Also assume the row can be stored on one base page and one extent page.
Since the row width is greater than the usable row size minus the reserved space,
SQLBase will be forced to store the row on a new page (this is because the existing
page will not contain sufficient freespace to store the row). Consequently, at the time
of the insert, both the base page and extent page will be allocated. If the first two
pages listed in the group’s free extent list are physically adjacent, then the base page
and extent page will be collocated. Otherwise, the base page and extent page will be
physically stored in separate blocks on the I/O device and two physical I/Os will be
required to retrieve the entire row.
One way that this latter situation can occur is if the group’s free extent list does not
contain two free pages and SQLBase must allocate additional pages to the group, or if
some pages have been de-allocated and placed back in the group’s free extent list.
Conversely, this latter situation cannot occur during a database reorganization since
no other users are connected to the database and thus are not competing for page
allocations. During a reorganization, SQLBase collocates all row pages with their
respective extent pages. However, it is possible that the row page and its extent page
are collocated but not stored in the same physical block. This situation would occur
when the base page was physically adjacent to the extent page but also crossed a
block boundary.
In any event, a database with too many allocated extent pages usually results in
decreased performance. A general design goal should be to minimize the number of
extent pages. To accomplish this, the DBA is encouraged to reorganize the database
frequently. In addition, if it is known that most of the rows in a given table will
significantly grow in width over time, then the CREATE TABLE’s PCTFREE
Advanced Topics Guide
Data pages
parameter should be adjusted accordingly. Further, if the average row width is close to
the usable page size, the database designer should consider splitting the table into two
Long varchar pages
SQLBase stores all long varchar column data in long varchar pages. Long varchar
pages contain only a page header, page footer and long varchar data. Long varchar
pages are chained to the base page by pointers in the base page slots. If a row contains
one or more non-null long varchar columns then its slot contains the set {column
identifier, column length required in the base page, first long varchar page physical
page number, last long varchar page physical page number, column size} for each
non-null long varchar column. Note that the column length required in the base page
is the size of the fields stored in the base page slot which for LONG VARCHAR
columns is always 12 bytes (consisting of the first long varchar page physical page
number, last long varchar page physical page number, column size). This page range
forms a linked list in the usual way, long varchar pages contain next and previous
page pointers in the page header (see Figure G).
E xtent P ages
Base R ow Pages
E xte n t
Pa ge
T yp e
Gr oup
E x te n t
Pa ge
T yp e
Grou p E x te n t
Pa ge
T y pe
Gro up
E xte n t
Pa ge
S lo t 2
S lo t 2
Slo t 2
LO N G V AR C H AR P ages
10 14
S lo t 2
R o w H e ad er
R o w S e r ia l N b r
1 10F r e d Sm ith
D a ta
L e n g th
U p d a te S e r ia l N b r
N br C ol um n s
C o l ID
2 7A tla n ta
Gr oup
1 015
101 5
Gro up
1 01 4
9 1 2 1014:10151611
1 s t : la s t p g
L O N G V A R C H A R D a ta
L O N G V A R C H A R D a ta
C o l ID & L e n g th
E n la rg e m e nt
Figure 8G - Unlike EXTENT pages, LONG VARCHAR pages are linked to their base
page by pointers in the slots. LONG VARCHAR pages are also linked together with
next and prior pointers. Please note in the enlargement of the base slot above that
displaced columns, those stored on EXTENT pages, are not referenced in the base slot
but LONG columns are referenced.
Advanced Topics Guide
Database Pages
The purpose of long varchar pages is to separate and isolate long varchar data from
columns of other data types so that the storage of this type of data does not negatively
affect the access of the other columns. Long varchar columns are used to store large
quantities of unstructured data such textual comments, bitmaps, video clips, and so
on. Consequently, one long varchar column, because of its size, will likely saturate
many pages. If these columns were stored in row pages, the result would be heavy use
of extent pages that would decrease performance to columns of all data types. By
storing long varchar columns on separate pages, access to non-long varchar columns
is unaffected.
On the other hand, a database access to a long varchar column as well as other
columns requires at least two physical I/Os (one for the row page and one for the long
varchar page). In addition, SQLBase stores only one column on long varchar pages. If
a column is defined as LONG or LONGVAR but its value is only one byte, an entire
long varchar page is used to store this one byte of data. Consequently, the LONG data
type should only be used when the values of the attribute are known to exceed 254
characters; otherwise, the VARCHAR or CHAR data type should be used.
Overflow pages
Overflow pages store rows when an overflow condition occurs in conjunction with a
clustered hashed table (see chapter 10). When a row hashes to a page which is already
full, SQLBase stores this row on an overflow page. There are two types of overflow
pages, designated overflow pages and dynamically created overflow pages. All
overflow pages, regardless of type, participate in the same linked list as base row
pages (see Figure C).
Designated overflow pages
When a CREATE CLUSTERED HASHED INDEX statement is executed, SQLBase
pre-allocates and creates row pages. The number of pages allocated depends on the
number of rows declared in the CREATE CLUSTERED HASHED INDEX statement
and the width of the table (see chapter 10). For every N number of row pages preallocated, SQLBase also pre-allocates a designated overflow page (see Figure H). If
Advanced Topics Guide
Data pages
the clustered hashed index is created on a single number field, N is 100; otherwise, N
is 10.
R o w s w h i c h o ve rflo w i n th is p a ge
ra n g e, a re s to red in the fol lo w in g
de s i gn at ed ov erflo w p ag e .
R ow s w h ic h o ve rfl o w i n th is p ag e
ra ng e, are s tore d in th e fo ll ow in g
de s ig n ate d ove rflo w pa g e .
D es ig na te d
O F P ag e
Figure 8H - Designated overflow pages are created for every N number of row pages
and participate in the same linked list as the base row pages.
Dynamically created overflow pages
Dynamically created overflow pages are created whenever both of the following
conditions are met:
A designated overflow page becomes full
SQLBase attempts to store another row in the designated overflow page
When this situation occurs, SQLBase will dynamically create another overflow page.
This new overflow page is then inserted into the linked list between the designated
overflow page and the next base row page (see Figure I). In this manner, SQLBase
will dynamically create as many overflow pages as is necessary to store the rows
which overflow in a given page range.
D yn a m ic al ly C re ate d
O v erf low P a ge
D e s ig na te d
O ve rflo w P a g e
Figure 8I - SQLBase dynamically creates as many additional overflow pages as is
necessary to store the rows which overflow in a given page range. These additional
Advanced Topics Guide
Database Pages
overflow pages are inserted in the linked list between the designated overflow page
and the next base row page.
SQLBase follows these steps when attempting to retrieve a row from a hashed table:
Hashes the key to obtain the target row page number;
Retrieves the target row page and searches the slots for the desired row;
If the desired row is not found on the target row page, SQLBase then retrieves the
designated overflow page (for that page range) and again searches the slots for the
desired row;
If the desired row is not found in the designated overflow page, SQLBase searches
in turn the dynamically created overflow pages.
Hence, too many overflow pages degrades performance. Care must be taken when
choosing a clustered hashed index key to ensure too many overflow pages will not be
required for any specific page range (see chapter 10).
Index pages
Index pages contain B +_tree indexes (refer to chapter 9, BTree Indexes, for a detailed
description of B +-trees). Index pages have the same basic organization of all other
page types. Each index page contains a page header, index page header, slot table,
slots and page footer. The page header, slot table and page footer is identical to data
pages (see Figure J).
Advanced Topics Guide
Index pages
P ag e H ea d e r
Ind ex Pa g e H ea d e r
P hys ic a l P ag e G r o up
L og ic a l
Pr e vi ou s
P ag e N o . T ype N um b e r P ag e N o . P a ge N o.
N e xt
Lo g
P a g e No . P oin ter
Sl o t
S iz e
F re e
S pa c e
In d e x
L ev el
Sl o t T ab le
E ntr y 1 En tr y 2 E nt ry 3
85 8
F r eesp ace
S lo t 3 ( 4 0 b y tes )
8 58
S lo t 1 (10 b yte s )
92 1
P a ge Fo o ter
L og Po i nte r
1 02 0 10 21
1 02 4
S l ot 2 (2 3 b y te s )
P a g e Nb r (4 by tes )
K ey Le ng th ( 1 by te)
S lo t N br ( 1 byt e)
P oi n te r Le ng th ( 1 byte)
89 8
P a g e P o in te r
S lo t O ve r he a d
K ey O ff s et ( 2 bytes )
7A tl an ta
A ttr 1
L en g th
5G u pta
Attr 2
L en gth
9 20
Figure 8J - Index pages have the same basic layout as row pages: each page contains
a page header, index page header, slot table, and multiple slots.
The index page header and slots within an index page differ from data pages. The
index page header contains the slot size, free space and the level in the index (see
Figure J). The slot contains the slot overhead, one page pointer and one or more key
attributes depending on whether the index is created on one or more table columns. In
the leaf pages, the page pointer points to a data page where the key values may be
found. Otherwise, the page pointer points to the next level in the index.
All index pages belong to the same group; however, non-leaf pages are not linked
together. Therefore, the next and prior page pointers in the page header are not used.
Leaf pages are linked together in the usual way and form what is known as the
sequence set.
Advanced Topics Guide
Database Pages
Control pages
Each SQLBase database contains a number of control pages for a variety of
housekeeping chores. Some of these control pages store various information about
the database as a whole. For example, the database control block is used to store
database parameters. Each SQLBase database also has one free extent list which is
used to keep track of un-allocated pages. Other control pages are used for groups and
tables. The group extent map is used to track pages allocated to a specific group.
SQLBase also has several types of control pages for each table defined in the
database. Each of these types of control pages is discussed in the following sections.
Database control block
The first page of every SQLBase database is referred to as the database control block,
or DCB. This page contains various database attributes such as:
SQLBase version number.
Recovery flag which indicates whether SQLBase shuts down gracefully or
When checkpoints were done.
Current log file.
Next log to backup.
Read-only flag which indicates whether the read-only isolation level is
Logging flag which indicates whether logging is enabled.
Log file pre-allocation flag which indicates whether log files are preallocated or not.
Log file size field, which is maintained through the SET LOGFILESIZE
command and indicates what size log file should be pre-allocated when using
A flag indicating whether the database is read-only, maintained through the
The extension size field, maintained through the SET EXTENSION
command, which specifies the disk space increment at which the database file
will grow.
The log file span limit, maintained through the SET SPANLIMIT command,
which specifies the maximum number of log files that one transaction may
be spread across.
Advanced Topics Guide
Control pages
The checkpoint time interval, maintained through the SET CHECKPOINT
command, which specifies the frequency in minutes between recovery
checkpoint operations.
When a user connects to a SQLBase database, SQLBase checks the version number in
the DCB to verify that a compatible version of SQLBase created the database.
Likewise, the recovery flag and current log file are used to determine whether
SQLBase should initiate crash recovery operations to rollback uncommitted
transactions (see Chapter 14, Logging and Recovery). The remaining fields stored in
the DCB are similarly used by SQLBase for various housekeeping functions.
The DCB is also used to store user parameters for the database. Any database specific
parameter set by the user and not stored in the sql.ini file is stored in the DCB.
Free extent list
As discussed previously, a SQLBase database file is a collection of pages. Database
pages are either allocated to a group or un-allocated. The free extent list is a set of
database control pages which maintain a record of which pages are un-allocated
(“free” for use). This list is physically stored as a B+Tree index (see Chapter 9, BTree
Indexes) where each node in the index includes the set {page number, number of
contiguous un-allocated pages}.
Group free page list
If a page is not in the free extent list, then it has been allocated to a group. For each
group, a list is maintained of its allocated pages. Like the free extent list, the group
extent map is physically stored as a BTree index where the node of the index contains
the set {page number, number of contiguous allocated pages}.
Group extent map
The Group Extent Map maps logical pages which comprise a group to physical
database pages.
Bitmap pages
SQLBase uses bitmap pages to track which row pages have freespace (remember the
row page header contains the amount of freespace. See Figure E). Each table within
the database has a set of bitmap pages for this purpose. Each of the 8192 bits (1024
bytes * 8 bits per byte) of the bitmap page (less overhead) maps to one row page.
When the bit is off, this signals the page has room for additional rows. When the bit is
on, the page does not have room for additional rows. There is one bitmapped page for
approximately every 8192 row pages allocated to the table’s group (a bitmap page
actually cannot track the freespace of 8192 pages since a portion of the bitmapped
page is used for overhead purposes, these details are ignored in this example). The
Advanced Topics Guide
Database Pages
first logical page within the group is a bitmap page and every 8192 pages thereafter is
a bitmap page. As SQLBase inserts, updates and deletes rows from a specific row
page, the freespace field in the row page header is adjusted accordingly. In addition,
when the page is 80% full, the corresponding bit in the corresponding bitmap page is
turned on. When allocated space drops to 50% or below, the bit is turned off.
Row count page
In addition to bitmap pages, each table defined in the database contains one row count
page. This page contains only one field, the number of rows in the table. This page is
updated whenever rows are inserted or deleted in the table. Consequently, this field is
frequently accessed and could become a bottleneck if it were not stored on a separate
Table data page
Each table in the database also has one table data page. This page is used to store:
Last row serial number assigned
Page number of last row page
Table size in the number of pages (hashed table only)
Number of overflow pages (hashed table only)
The first two fields are accessed by SQLBase whenever new rows are inserted into the
database. The last assigned serial number is used to determine the serial number for
the new row. When a new row is inserted into the table, SQLBase uses the last row
page number to quickly locate the last page. The table data page contains the latter
two fields only if the table is hashed (is the subject of a clustered hashed index).
Page allocation and garbage collection
The process of allocating new pages and reclaiming space in existing pages when
rows are deleted is know as page allocation and garbage collection. Each of these
processes is discussed in the following sections.
Page allocation
When SQLBase receives a command to insert a row into a table, SQLBase accesses
the corresponding table data page in order to retrieve the page number of the last row
page in the linked list. SQLBase then retrieves this page and checks the freespace
field in the row page header to determine if room exists to store this specific row. If
so, the new row is inserted into the last page in the linked list and the freespace field
in the row page header and corresponding bitmap page are updated accordingly.
Advanced Topics Guide
Page allocation and garbage collection
If the last page in the linked list does not contain sufficient freespace to accommodate
the new row, SQLBase begins searching each bitmap page in turn, starting with the
last bitmap page accessed and wrapping around to the first bitmap page if necessary,
for an existing row page with its bit turned off. If one is found, SQLBase retrieves this
page and checks the freespace remaining in the row. If sufficient space exists, then the
row is stored on the page. If sufficient space does not exist, SQLBase continues its
search of the bitmap pages. It is also possible for SQLBase to encounter a fragmented
page; that is, a page with enough total free space to store the incoming data, but not
enough contiguous free space. In this case, garbage collection is performed to
compact the page and coalesce all of the free space into one chunk (see “Garbage
collection” below for details).
If all bitmap pages are searched and no page is found with sufficient freespace,
SQLBase then checks the free extent list. If un-allocated pages exist in the free extent
list, SQLBase allocates a set of pages, known as an extent, to the table’s group by
removing them from the free extent list and adding them to the group extent map. At
this time, SQLBase also updates the appropriate bitmap page and/or creates new
bitmap pages if necessary. Groups grow in intervals of 10% or five pages, whichever
is larger. For example, if the group currently consist of 100 pages, 10 new pages are
allocated to the group. On the other hand, if the group only contained 40 pages, then
the extent would consist of five pages.
If no unallocated pages exist in the free extent list, SQLBase performs the following
Obtains additional space from the file system. The SQLBase database file is not
fixed to a certain size; it grows as additional space is required. When a SQLBase
database requires additional storage space, additional pages are added to the
database file. This process of adding additional pages to the database file is
resource intensive, increasing response time for currently active transactions.
Consequently, pages are not added one at a time but in groups of pages called
extents. The extent size may be set by the user with the SET EXTENSION
command. For non-partitioned databases, the default extent size is 100 pages.
Adds extent pages to the free extent list.
Removes pages from the free extent list and added to the group extent map.
Updates the row and freespace field stored in the new page and the corresponding
bit in the bitmap page.
Advanced Topics Guide
Database Pages
Garbage collection
When a row is deleted from a page:
The utilization bit in the slot table is turned off, indicating that the slot is
available for re-use.
The freespace value in the row page header is increased by the row size.
If freespace is now greater than 50% of the usable page size minus reserved
space, the corresponding bit in the bitmap control page is turned off.
At the time of the deletion, the slot data is not removed and the page is not
compacted. Slots are reshuffled to free up space when an insert or update (which
increases the size of the subject row) occurs. In this way, garbage collection is only
performed when needed.
Advanced Topics Guide
Advanced Topics Guide
Chapter 9
B-Tree Indexes
In this chapter we:
Introduce the concept of tree structured indexes, particularly the B-Tree
index, a variation of which is used by SQLBase.
Discuss the general algorithms used to maintain B-Tree indexes.
Introduce the SQLBase variation of B-Tree indexes, the B+-Tree index, by
contrasting it to the B-Tree index.
Explain the variations between the algorithms used by SQLBase to maintain
B+-Trees and the basic algorithms already covered.
Discuss the unique considerations that apply to SQLBase indexes.
Advanced Topics Guide
B-Tree Indexes
B-Tree indexes have become the de facto standard for random access to records (or
rows) within a database file. All major RDBMS products use one variety of B-Trees
or another for general purpose indexing. SQLBase uses the B+-Tree variation first
proposed by Knuth in 1973.
Concerning the history of B-Tree indexes, Douglas Comer documents:
The B-tree has a short but important history. In the late 1960s computer
manufacturers and independent research groups competitively
developed general purpose file systems and so-called “access methods”
for their machines. At Sperry Univac Corporation H. Chiat, M.
Schwartz, and others developed and implemented a system which
carried out insert and find operations in a manner related to the Bl-tree
method. . . . Independently, B. Cole, M. Kaufman, S. Radcliffe, and
others developed a similar system at Control Data Corporation . . .
R. Bayer and E. McCreight, then at Boeing Scientific Research Labs,
proposed an external index mechanism with relatively low cost for most
[file] operations; they called it a B-tree. The origins of “B-tree” has
never been explained by the authors. As we shall see, “balanced,”
“broad,” or “bushy” might apply. Others have suggested that the “B”
stands for Boeing. Because of his contribution, however, it seems
appropriate to think of B-trees as “Bayer”-trees.
B+-Tree Before presenting the details of the B +-Tree variation used by SQLBase, it is
necessary for the reader to understand simple binary search trees, multiway trees and
B-Trees in general. This chapter presents a general introduction to binary search
trees, multiway trees, and B-Trees. Following this general introduction, B-Trees are
presented along with the SQLBase implementation.
Advanced Topics Guide
Binary search trees
Binary search trees
When traversing a binary search tree, the path chosen is determined by comparing the
target key to the key value stored at each node. If the target key is less than the node
key, the left branch is followed. Otherwise, the right branch is chosen. Figure A shows
part of such a search tree and the path taken for the target key “15.”
Figure 9A - This figure shows a portion of a binary search operation. The path
followed for the target key ‘15’ is illustrated with dotted lines.
Each node in a binary search tree contains exactly one key and has a branching factor
of two: every node contains a maximum of two children. Thus, binary search trees are
“tall” trees. The minimum height of a tree containing N records is [log2N]+1. For
example, a tree with only 100 records will have at least seven levels.
When trees are used for index structures, “short” trees are preferable since they
typically exhibit better performance characteristics. This is because typically one
physical I/O is required for every level traversed in the index, and in order to speed
database access times, physical I/O must be kept to a minimum.
Note also that in index structures, a record in a node contains the set {Key Value,
Pointer}. The pointer is the physical address in the main file where the key value can
be found. Figure A above contains only the key values in the nodes of the tree in order
to simplify the illustration. In the remainder of this chapter we continue to follow this
same practice.
Advanced Topics Guide
B-Tree Indexes
Multiway trees
Multiway trees are a generalization of the binary search tree data structure. With
multiway trees, each node contains N number of keys. For example, consider Figure
B which illustrates a multiway tree with a node size of 2.
n od e a
nod e b
no d e c
no de d
Figure 9B - This figure shows a search tree with two keys and three branches per
node. The path followed for the target key “15” is illustrated with dotted lines.
Consider the path followed to locate the target key “15.” First, node a, or the root of
the tree, is examined. The target key of “15” is less than “42”, the first key in node a,
so the left most branch is followed. Next, node b is examined. The target key is
greater than “7”, the first key in node b, so the target key is compared to the second
key in node b. Since “15” is less than “24” the middle branch is followed. This
algorithm is repeated until either an exact match is found or a leaf node is
encountered (a leaf node is any node which is the furthest from the root, in other
words, a leaf node has no child). If a leaf node is examined with no exact match
found, then the target key is not contained within the index.
Multiway trees are generally “shorter” than binary trees since their branching factor
is greater. However, complex algorithms are required to keep multiway trees
“balanced.” In a balanced tree, all leaf nodes appear at the same level, and the tree has
the same height regardless of which leaf node is measured. A balanced tree is
desirable since it results in uniform performance.
B-Trees are balanced multiway trees with the following additional key property: each
node need not contain exactly N number of keys. Nodes are allowed to “grow” from
half full to full. This additional property greatly simplifies the balancing algorithms
and significantly reduces the cost. B-Trees are a classic example of the disk space
versus performance trade off. Often, as in the case of B-Trees, additional disk space
can significantly increase performance. The rationale used in these cases is that,
Advanced Topics Guide
presumably, additional disk space is relatively inexpensive and justified for the
desired increase in performance.
A B-Tree of order d can be defined formally as a tree with the following properties:
Intermediate nodes (meaning nodes which are not leaves or the root) have:
At least d records (keys) and no more than 2d records (keys);
At least d+1 children (pointers) and no more than 2d+1.
The root, unless the tree has only one node, has at least two children.
All leaf nodes appear at the same level, that is, are the same distance from the
root (meaning the tree is balanced).
The order of a B-Tree
The order of a B-Tree describes the “bushiness” of the tree: the number of records
stored in each node and consequently the number of child nodes. The root and leaf
nodes are exceptions but all other nodes in a B-Tree of order d have between d and 2d
records and between d+1 and 2d+1 children. For example, a regular node (meaning
one that is not a leaf or root node) in B-Tree of order 10 has at least 11 and not more
than 21 children. The lower bound ensures that each node is at least half full;
therefore, disk space is not unnecessarily wasted. The upper bound allows the B-Tree
to be mapped to fixed length records in an index file. The order also effects the
performance of the B-Tree index. If the B-Tree gets to “tall” (has too many levels),
performance will suffer. Likewise, if the order gets too “wide” (has too many records
per node), performance will also decline.
Sequencing within a node
In addition to the data structure properties defined in the formal definition above, each
node of a B-Tree index also exhibits some sequencing properties. Specifically:
Within a node of K records, records may be numbered R1, R2, R 3, ..., R K and
pointers to child nodes may be numbered as P0, P1, P2, ..., P K. Figure C
illustrates a typical node using these number semantics.
Figure 9C- This figure illustrates a typical node in a BTree index. Note that each node
has K records but K+1 pointers to child nodes.
Records in the child node pointed to by P 0 have key values less than the key
of R1. Records in the child node pointed to by PK have key values greater than
Advanced Topics Guide
B-Tree Indexes
the key of RK . Records in the child node pointed to by Pi (where 0 < i < K)
have key values greater than the key of R i and less than the key of R i+1.
Balancing B-Trees
The popularity of B-Tree indexes results from their performance characteristics
which is derived from their balancing algorithms. Records are usually inserted into
and deleted from a database (and therefore an index) in an arbitrary fashion. Random
insertions and deletions can cause trees to become unbalanced (see Figure D).
Leaf Nodes
Figure 9D - Unbalanced trees have some short and some long paths from the root to
the leaf nodes.
With an unbalanced tree, some leaf nodes are further from the root than others;
consequently, path lengths and ultimately retrieval times vary. Some key values may
be located by only visiting a few nodes; other key values may require accessing a
great many nodes. Balanced trees, as mentioned previously, are trees whose leaves
are all the same distance from the root. This characteristic results in uniform search
paths and thus consistent retrieval times. However, in order to maintain a balanced
tree through random insertions and deletions special balancing algorithms must be
B-Tree balancing algorithms are specially derived to minimize balancing costs. First,
B-Tree balancing algorithms confine node changes to a single path from the root to a
single leaf. In addition, by utilizing extra storage (where some nodes are not 100%
full), balancing costs are also lowered. Since nodes are sometimes not full at the time
of an insert, the new record can simply be inserted into the node (therefore no
balancing action must take place). Similarly, when deletes occur, sometimes the
target node contains more than the minimum number of records so, again, the record
is simply deleted. Only when the node contains 2d records (inserts) or d records
(deletes) are balancing steps required.
Advanced Topics Guide
Figures E, F and G graphically illustrate the classic B-Tree insertion algorithm. Figure
E is the base tree before any insertions. To insert the key “Q” into the B-Tree of
Figure E, the find algorithm is followed to locate the leaf node where the new key
belongs. If we follow this algorithm we discover that node 7 is the node where the
new key belongs. Since this node has only two keys, the new key value of “Q” can
simply be inserted in the third record.
nod e 1
node 3
node 2
n ode 6
node 5
node 4
no de 7
n ode 8
Figure 9E - This is a B-Tree of order 2, meaning each node has between two and four
Figure F is the same tree as Figure E after the key of “Q” has been inserted. This
insertion represents the usual case: no balancing steps are required because the node
is not full. B-Trees trade extra storage space (when some nodes are not full) for added
node 1
node 2
node 4
node 3
node 6
node 5
node 8
node 7
Figure 9F - This is the same B-Tree as Figure E after the key “Q” is inserted.
Let’s examine a more complex example. Suppose we insert the key value of “U” into
the same tree. In this case, “U” belongs in node 8; however, this node already has four
keys. In this situation, a “split” must occur. To split a node, the keys with the lowest
values are placed in one node and the keys with the highest values are stored in a new
node. In addition, the “middle” key is propagated to the parent node and serves as a
separator. Figure G contains the same tree as Figure F after “U” is inserted. In this
example, the parent node contains free records for additional keys. If the “middle”
key cannot be placed in the parent because the parent node is also full, then the parent
Advanced Topics Guide
B-Tree Indexes
node is also split. In the worst case, this splitting occurs all the way to the root and the
B-Tree grows by one level.
node 1
n ode 2
n ode 4
nod e 3
node 5
nod e 6
nod e 7
node 8
node 9
Figure 9G - A representation of the same B-Tree as Figure E after the key “U” is
The delete algorithm, like the insert, begins with the find or retrieval operation in
order to locate the node which contains the target key. The target node may be found
either at the leaf level or at any other level within the tree.
If the target node is a leaf node, the key is simply removed. Since the majority of the
keys are stored at the leaf level (because there are many more leaf nodes than non-leaf
nodes), most of the time the target node will be found at the leaf level.
If the target node is not a leaf node then the key cannot simply be removed. The key
must be replaced with a new “partitioning” key so that future find operations are
guided properly to the child nodes. The general procedure is to replace the target key
with the next highest key value, or what is known as the “successor” key. The
successor key will always be found at the leaf level by traversing the right-most
branch of the tree. Once the successor key is located, it is simply moved from the leaf
node to the target node. In other words, the successor key replaces the target key (the
one to be deleted).
In either case, whether the target node is a leaf or not, the above step causes one leaf
node to shrink in size by 1 record. When this occurs, the affected leaf node may now
have too few keys, meaning it is less than half full. This situation is termed an
underflow. Two techniques are then used to deal with the underflow condition:
redistribution and concatenation.
If either one of the adjacent siblings of the affected leaf node contain excess records
(where the nodes are more than half full) then the redistribution technique may be
utilized. Redistribution simply moves keys among one sibling node, the affected node
and the parent such that the underflow condition is remedied.
Advanced Topics Guide
On the other hand, if neither of the affected leaf node’s adjacent nodes have excess
records, then the concatenation technique must be applied. Concatenation involves
merging nodes and is, in effect, the opposite of the splitting operation used in the
insertion algorithm. Like splits, concatenations may propagate all the way to the root.
However, in this case, if the propagation travels to the root, the tree shrinks by one
level (rather than growing by one level as with splits).
Detailed examples of the deletion algorithm may be found in Comer, Smith and
Barnes, Knuth, as well as any introduction text to index organizations. Wirt provides
sample routines programmed in Pascal.
B-Tree operation costs
Knuth presents a detailed analysis of B-Tree retrieval, insertion and deletion costs. In
the following sections we present a simplified model of B-Tree operational costs.
Retrieval costs
If we assume each node in a B-Tree is mapped to a block on a physical I/O device,
then each node visited in a find operation requires one physical I/O. Consequently,
retrieval cost is proportionate to the number of nodes visited to locate the target node.
The number of nodes visited during the find operation is directly related to the height
of the B-Tree since most nodes will exist at the leaf level. In other words, most of the
time, the find operation will traverse the tree all the way to the leaves before the target
key is located. Hence, the number of physical I/Os required to locate the target key is
approximately equal to the height of the tree. Knuth and others have shown that for a
B-Tree of order d with n total keys:
height ≤ log  ------------
Thus, the cost of the find operation grows as the logarithm of the total number of
keys. The following table shows how efficient B-Trees are in terms of I/O due to their
logarithmic cost function, since the height of the index is generally equivalent to the I/
Os required to locate a leaf entry:
Nbr Keys =>
Node Size
Advanced Topics Guide
B-Tree Indexes
Nbr Keys =>
Node Size
Even with 1 million keys, one can locate a target key within a B-Tree of order 50 in
only four physical I/Os. In fact, Knuth and others show that this cost can be improved
upon in practice by some simple implementation adjustments. The average cost is
somewhat less than three physical I/Os.
Insertion and deletion costs
Both the insertion and deletion algorithms begin with the find operation.
Consequently, insertion and deletion cost is a superset of retrieval costs. In addition,
when a key is inserted or deleted into the tree, in the worst case splitting or
concatenation continues up the tree to the root. Therefore, maximum physical I/O for
an insertion or deletion is double that of the find operation.
As we have shown, B-Trees offer near optimal performance for random access to one
record. However, sequential processing using a standard B-Tree is both inefficient
and difficult. After locating the first record using the find operation, to locate the next
sequential record may require traversing from the leaf to the root and back down to
another leaf. Thus, in the worst case the cost of the next operation is:
MaxCostOfNext = log n
For example, consider the following SQL SELECT statement:
This SQL statement retrieves only one row; consequently, a standard B-Tree index
provides efficient random access to this one row. On the other hand, consider another
SQL SELECT statement:
Unlike the first SQL statement, this SQL statement retrieves all customers in
sequence by the ID attribute. To satisfy this query, one find operation would be issued
to locate the first customer followed by a find next operation for every customer
stored in the customer table. Since the customer table could contain millions of rows
and each find might require MaxCostOfNext (as defined above) physical I/Os, the
cost of processing this query using the standard B-Tree would be quite large.
Advanced Topics Guide
B+-Trees are derived from the standard B-Trees. They add efficient sequential
processing capabilities to the standard B-Tree yet retain the excellent random access
Before defining B +-Trees, it is necessary to define two terms:
“Index nodes” refers to all nodes not at the leaf level, meaning the root and
all subsequent levels excluding the leaf level;
“leaf Nodes” or “leaves” are all nodes at the leaf level.
B+-Trees differ from B-Trees
in that (see Figure H):
All pointers to the main file are removed from the index nodes;
All key values are stored in leaves regardless of whether they exist in index
Leaf nodes are chained together into a linked list known as the sequence set.
Index N odes
Figure 9H - A B+-Tree is composed of index nodes and leaf nodes that are linked
together in a linked list known as a sequence set.
These structural differences result in several key implications:
All find operations must travel to the leaf level to obtain the pointer to the
main file; thus, the index nodes serve as a “road map” to the appropriate entry
point to the sequence set.
Since pointers to the main file are removed from the records in the index
nodes, more records can be stored in the index nodes; thus, the performance
is on a par with that of B-Trees.
Sequence set provides an efficient mechanism for sequential processing: at
most the cost of the next operation is one I/O.
Advanced Topics Guide
B-Tree Indexes
Prefix B+-Trees
Prefix B +-Trees are a slight variation of B+-Trees. Since the index nodes only serve as
a road map to the leaves and all operations must travel to the leaf nodes, the index
nodes do not have to store the full key values. The index nodes must store only
enough of the key value to distinguish one branch from another and direct the find
operation to the appropriate leaf (see Figure I).
gamm a
alp ha
gam m a
Figure 9I - The tree fragment on the left contains the full key in the index node. The
tree fragment on the right contains only the “minimum distinguishing key.”
Prefix B+ -Trees store the “minimum distinguishing key” in the index nodes. As a
result, many more index records may be packed into the index nodes thus increasing
the performance of the B +-Tree by decreasing the height of the tree.
SQLBase implementation
SQLBase B-Tree indexes are based on the B-Tree variant described above as the
prefix B+-Trees. This B-Tree variant provides excellent random and sequential
performance. Various aspects concerning SQLBase’s implementation of prefix B +Trees are discussed below.
Index node size
The performance of a B-Tree is, as we discussed above, largely dependent on the
height of the tree. As we also discussed, tree height is a function of node size (number
of index records per node) and the number of rows in the table.
However, since SQLBase only stores the minimum distinguishing key in non-leaf
nodes, node size is difficult to predict for non-leaf nodes. This is because the savings
in secondary storage achieved through only storing the minimum distinguishing key
is hard to predict because it is dependent upon the actual values. As a general rule,
however, a 2:1 (or better) compression ratio is generally achieved. In other words, the
node size of non-leaf nodes may be estimated as two times the size of leaf nodes.
SQLBase index nodes are based on a database page (see Chapter 8, Database Pages).
For DOS, OS/2, Windows NT, and NetWare, the page size is 1024 (for UNIX
Advanced Topics Guide
SQLBase implementation
implementations of SQLBase the pages size is 2048 bytes). Index page overhead is 74
bytes. The node size of leaf nodes is calculated as:
UsableInd exPageSize
Node Size = ---------------------------------------------------------IndexEn tryLength
100 – Perc entFree
UsableIndexPageSize = ( 1024 – 74 ) × ----------------------------------------------- 
IndexEntryLength = 9 + NbrColumns + KeyLeng th
For example, if an index contains three attributes, each of which contains, on average,
20 characters, then the B-Tree node size is:
NodeS ize =
( 100 – 5 )
950 ---------------------100
-------------------------------9 + 3 + 60
950 ( 0.95 )
= 12
Hence, in this example, the node size of leaf nodes is twelve and one would generally
expect the node size for non-leaf nodes to exceed 24, in this case.
Performance considerations
In an ideal situation, a database designer could set up an index and then rarely have to
perform any maintenance on it. However, index performance degrades over time due
to a number of factors, foremost of which is the pattern in which insertions and
deletions occur. When this happens, index reorganization becomes necessary. This
reorganization can take one of two possible forms:
Reorganization of the entire database will rebuild all indexes within the
database as a subset of the process. This is overkill unless the entire database
has reached the point where total reorganization is required.
Dropping and then re-creating an index will cause the index to be built fresh.
This will result in an optimal, balanced tree structure for the index, and is the
simplest and quickest way to return a degraded index to peak performance.
Deletions can cause balancing problems in the tree structure and fragmentation of the
index blocks. SQLBase does not perform redistributions or concatenations when
underflows occur in delete processing, based on the philosophy that these operations
would take too much time during regular transaction execution. Rather, it is left to the
DBA to rebuild the index when the accumulated underflows are sufficient to impact
Insertions are not always spread uniformly across the index tree structure. When
many insertions occur within a small range of symbolic key values within the tree, the
freespace within this area is rapidly used up. Once this occurs, there can be frequent
splits within this portion of the tree, sometimes reaching all the way back to the root.
While these operations are taking place, many exclusive locks will be held throughout
Advanced Topics Guide
B-Tree Indexes
the index structure for the duration of the split operation. The locks may cause
contention problems for other transactions wishing to use the index structure during
this time, see chapter 13, Concurrency Control, for more details on the locking
implications of B-Tree indexes in SQLBase.
This imbalance is the primary reason why indexes placed on columns whose values
are incremental in nature, such as a timestamp, can cause performance problems
within an application. These indexes grow only along one edge. As rows are inserted
along this edge splits occur frequently, so that transactions updating this index find
themselves competing for locks within a small portion of the overall index structure.
One possible solution to this problem is to add columns to the index in order to make
the insertion pattern more evenly distributed throughout the tree.
Setting freespace for an index
Since the effect of repetitive insertion activity into an index can cause splits to occur,
which in turn can cause processing overhead and locking contention, DBAs may wish
to minimize the possibility of excessive splits occurring within indexes that are the
subject of high transaction activity. The PCTFREE specification of the CREATE
INDEX statement allows DBAs to realize this goal. Through correct calculation of
the PCTFREE figure, split activity in an index may be minimized for some period of
time following the creation or rebuilding of an index. How long this time period is
depends on both the rate of insertion activity against the index, and the pattern of the
symbolic keys being introduced into the existing index structure.
When creating an index, the database designer can choose a PCTFREE value to
specify in the CREATE INDEX statement. This value determines the amount of free
space that will be left in the index nodes during the index’s initial creation. This value
is fairly easy to calculate for an index where the insertion activity can be expected to
be evenly distributed throughout the index structure, by utilizing the following
Calculating PCTFREE for an index
Either estimate or obtain the current number of rows in the table. This can be
obtained directly through the sqlgnr function.
Estimate the number of rows that will be in the table at the point in time in the
future when you plan on rebuilding the index.
Calculate the PCTFREE parameter according to the following formula:
( CurrentRows – FutureRows )
PCTFREE = ------------------------------------------------------------------------------ × 100
If the value calculated above is greater than 50%, consider rebuilding the index
more frequently to lower the PCTFREE value.
Advanced Topics Guide
SQLBase implementation
This procedure can also serve as a starting point for an index that is known to have a
skewed insertion pattern. After obtaining a PCTFREE value from this procedure,
either increase the value by a factor of between 1.5 to 3, or simply plan on rebuilding
the index more frequently than the basis of your future row estimate would indicate.
The more skewed the insertion pattern, the more dramatic the adjustment should be to
ward off potential performance problems.
B-Tree locking
This topic is covered extensively in Chapter 13, Concurrency Control. To summarize,
indexes can often become points of lock contention when they are placed on an active
table, and should therefore be used sparingly in these cases. When many transactions
are being executed simultaneously, and rapid response time is critical, any insertion or
deletion activity in index structures will have a significant adverse impact.
Non-unique indexes
Indexes which lack the “UNIQUE” clause on the CREATE INDEX statement can
contain more than one row with identical symbolic key values. SQLBase handles
these entries as if they were unique anyway, in that each entry exists independently of
any other entry. This is in contrast to some DBMSs which build synonym chains for
duplicate entries, and then incur significant overhead while maintaining these long
synonym chains. For this reason, SQLBase can handle any number of duplicates
within an index without incurring any significant performance degradation, and in
fact some space savings will be realized due to the minimum distinguishing key
One potential drawback to non-unique indexes, though, is their lower selectivity
factor. This is a statistic that is kept for each index by SQLBase, and is used by the
query optimizer (covered in detail in Chapters 15 through 18) when selecting indexes
to use when building an access path to satisfy a query’s data requirements. If an index
has many duplicates as a percentage of its total rows, then there is a good chance that
the index may rarely, if ever, be selected for use by the optimizer. Such indexes are of
no benefit whatsoever to the system, and should be deleted.
An alternative action that will increase the selectivity factor of the index is to add one
or more additional columns to the low order end of the index which will enhance its
uniqueness. In fact, one technique that is sometimes used is to add the primary key of
the table to the low order of an index, which can then be created as a unique index.
Note, though, that this technique will not effect the optimizer’s selection of the index
unless these additional columns can actually be specified in the WHERE clause of the
SQL referencing the table. This is because the optimizer considers the selectivity
factor only for those index columns that are actually specified in the statement being
prepared for execution, as opposed to the entire composite index key.
Advanced Topics Guide
Advanced Topics Guide
Chapter 10
Hashed Indexes
In this chapter we:
Introduce the concept of hashing, and contrast its effect on the location of
rows within a table’s storage space with that of non-hashed tables.
Analyze the theory behind the performance benefits that can be expected
when hashing a table.
Explain the details of SQLBase’s implementation of hashing including the
steps you need to take to ensure that your hashed tables will meet your
performance expectations.
Advanced Topics Guide
Hashed Indexes
Hash table overview
SQLBase offers an alternative to the B+Tree index structure covered in the previous
chapter. This alternative is the clustered hashed index, which for the remainder of this
chapter will simply be called hashing. A database table that has a clustered hashed
index placed on it is called a hashed table.
Row location in hashed tables
Hashing is an entirely different approach from the conventional B+Tree index in that
the basic allocation algorithm for table rows is fundamentally changed. In other
words, the main effect of a clustered hashed index is on the subject table itself,
whereas the main effect of a B+Tree index is to allocate disk storage and build a
B+Tree representation of the subject column’s symbolic key values, which contains
pointers into the subject table. B+Tree indexes do not alter the subject table, whereas
hash indexes do. This is the reason that a hash index must be created prior to any rows
being inserted into the table, if the subject table of the CREATE CLUSTERED
HASHED INDEX statement is not empty an error will be returned.
It is important to remember that a clustered hashed index has no physical index object
as the B+Tree index does. The effect is entirely on the allocation scheme used for the
subject table. Also, the term “clustered” is something of a misnomer, since there is no
sequential order to the storage of the rows that corresponds to the collating sequence
of the symbolic key. The “clustered” aspect is that rows which have equal symbolic
keys will be stored on the same pages, as contrasting with non-hashed table row
allocation which is governed strictly by the timing of the insert commands (roughly
equivalent to being clustered chronologically by insertion time).
Performance benefits of hashed tables
Hashing is a file organization technique which provides very fast access to rows in the
file. A properly implemented hashed table will allow access to almost all rows with a
single read operation. In order for the hash access technique to be utilized, the
SELECT ... WHERE clause must specify an equality predicate on the hash field of
the table. This hash field is the column, or columns, named as the subject of the
CREATE CLUSTERED HASHED INDEX statement. The hash field (or the
concatenation of columns which comprise the hash field) is also commonly called the
hash key. Example 1 illustrates the importance of the hashing method to a database
system in saving disk accesses (I/Os) when searching for a record.
Example 1: Suppose there is a file with 50,000 records stored on a disk with
block size 1,024 bytes. We assume that records are of fixed size with
a record length of 100 bytes. Hence there are 1024/100 = 10
records in each block, where x is a floor of x, which truncates the
fraction part of x (for example, 3.1 = 3). The number of blocks
Advanced Topics Guide
Hash table overview
needed for the file is 50,000/10 = 5,000 blocks, where x is a ceiling of x, which rounds up the fraction part of x (for example, 3.1
= 4). Access of the file through a BTree index would require one I/
O for the leaf page, one for the actual data block, plus another I/O
for each level in the BTree structure (probably at least another two
I/Os, although would probably be improved through caching). This
would make a total of approximately four I/Os for the BTree access.
However, hashing makes it possible to find a record with very few
disk accesses, typically one disk I/O. This means that the hash access
would require only one-fourth of the I/O time that the BTree access
Potential pitfalls of hashing
This improvement in performance does not come free, of course. The following
drawbacks to hashing should be considered before selecting this access technique:
The hash function must work well with the selected hash key, eliminating bias
to minimize overflow. Fortunately, SQLBase has an excellent hash algorithm
built-in which works well for most keys.
Rows of hashed tables cannot be accessed sequentially in the sequence of the
primary key without undergoing an intermediate sort. If sequential access is
often required, then a B+-Tree index on the primary key would have to be
created to provide an efficient access technique that would bypass the need to
perform a sort operation.
Hash keys have no generic capabilities, therefore the LIKE, less than (<),
greater than (>), BETWEEN, or similar predicates cannot be used with
hashed access. The only predicates that can be used are the equal (=), for
single value retrieval, or the IN list, for multiple value equality conditions.
Also, for multi-column hash keys, all columns must be specified, each with
one of the two allowable predicates. Again, these capabilities can be provided
on a hashed table by creating a B+-Tree index on the table for the desired
Space for the hashed table will be pre-allocated at table creation time,
specifically when the CREATE CLUSTERED HASHED INDEX statement
is executed. This space allocation is one of the parameters used by the hash
function when computing row locations. For this reason, changing the space
allocation requires the table to be unloaded, dropped and redefined, and then
reloaded. Because this may be a very time consuming process, tables which
experience wide unpredictable swings in row volume are generally not well
suited for hash treatment. Tables which maintain a consistent size or a slow
predictable growth are the best hash candidates.
Advanced Topics Guide
Hashed Indexes
Hashing theory
We will now examine the characteristics of the hashing storage technique in general.
We will cover the hash function, which performs the transformation of symbolic hash
key values into storage addresses, the concept of bucket hashing, where more than
one record is held at the same hash address, and provide a detailed analysis of the key
performance criteria used to evaluate the success of a hashed file implementation.
These criteria all relate to a hashed file’s disk space allocation, and the probabilities
of the hashing algorithm being able to accomplish its objective of storing and
retrieving records with no more than one I/O.
The hash function
The idea of hashing is to utilize a hash function, H, which is applied to the hash key
and yields the address of the row:
H( HashKeyValue ) ⇒ RecordAddress
The purpose of a hash function is to transform the hash key (which may be a
concatenation of multiple columns) to a page address that lies within the range of the
pre-allocated file. The algorithm that the function uses is fairly arbitrary, so long as
the final result is always within the range of the file. It is also desirable that the
transformation be uniform, that is, there be no bias toward any particular address. If
there is bias, rows hashing to the “favored” addresses are likely to overflow their
target pages and take longer to store and retrieve (from overflow pages). The details
of a simple hypothetical hash function are shown in the next example.
Example 2: Consider a file where the hash field is a person's name. Suppose we
want to reserve space for 1,000 records, which implies that the result
of the function must be in the range of 1 to 1000. The hypothetical
hash function H can be defined as taking two numbers from the
ASCII representations of the first two characters of the name (which
is the hash key), multiplying them together (in an attempt to eliminate bias), and dividing the result by 1,000 and taking the remainder
for the hashing address (which insures the result will be in the proper range). Table 1 shows how three names would produce three
hashing addresses. Note that although the names are listed in alphabetical order, the addresses are not in this order, but appear to be random.
66 , 79
83 , 84
Advanced Topics Guide
Hashing theory
84 , 69
Bucket hashing
In practice, the I/O system retrieves a group of records from a disk, instead of a single
record, because records are grouped together in a block. We can extend the idea of a
record address in a file to the address of a block containing a set of records, and call
each block a hash bucket. In a bucket scheme, the hashing function generates a bucket
number instead of a record address as the hashing address for a given hash field. The
bucket number can usually be directly transformed into a relative block number in a
hashed file. When this relative block number is added to the starting disk address for
the file, any block in the file may be accessed directly via its disk address. Once the
bucket’s disk address is resolved, we need only a single block access to retrieve the
required record. The cost to search for the record within the block is minimal since it
can be carried out in a main memory buffer.
Collisions and overflows
A collision occurs during an insertion when two records are hashed to the same
address. Referring to our hypothetical hash function in example 1, suppose there is a
record in the file with the name ETHAN. Since the name ETHAN starts with the same
two characters as the name TED, they produce the same hashing address 796. There is
a collision between the record for TED and the one for ETHAN. One way to resolve
the problem of collisions is to find a hashing function that prevents collisions from
happening at all. Such a hashing function is called a perfect hashing. However,
perfect hashing functions are rare. Knuth showed that in the set of all possible
mappings of a 31-record set to a table with 41 locations, only about one out of 107 is
the perfect function. Therefore it is not practical to find a perfect hashing function.
Instead, a better strategy is to focus on determining an efficient technique to handle
collisions when they occur.
In bucket hashing, a collision occurs when two records hash to the same bucket.
Collisions are not serious if the bucket is not full, since records can be stored in
available spaces in the bucket. However, overflow arises when a new record is hashed
to a bucket which is already full. How SQLBase handles collisions in its bucket
hashing algorithm will be examined in a later section of this chapter.
Advanced Topics Guide
Hashed Indexes
Collision avoidance techniques
Two techniques are available to the database designer in order to reduce the
possibility of collisions in a hashed table. The first technique involves selecting a
hash key which will allow for the even distribution of rows across the allocated space
using the selected hash function. The other technique is to allocate a greater amount
of space than is needed in order to allow the hash function a greater page range of
allowable results to operate across. Application of these techniques is imperative if
the desired performance benefits associated with hashing are to be achieved.
Hash key selection
The first technique entails choosing the hash field carefully. A badly chosen key leads
to poor hashing results. An inappropriately selected hash key often has one of the
following characteristics:
Values of the key are not evenly distributed. In general, uniqueness is a
desirable characteristic for a hash key (one significant exception to this rule
is described in Chapter 6, Refining the Physical Design). Suppose there is an
employee table with each record consisting of attributes such as employee
social security number (SSN), name, and department code, among others.
The primary key for the employee is the 9-digit long SSN. The department
code is 2-digits long. The employee’s name is not a good hash key candidate,
since there is an uneven distribution of names in most organizations (many
“Smith” and “Jones” occurrences). Also, the department code may not be a
good candidate either, since many employee records could be clustered
together in departments which have very large staffs. The best key for this
case is the employee SSN, which is unique. Serially created numeric fields
that contain arbitrary values are often found to be good hash key candidates.
The key does not match well to the hash function. Suppose that the employee
ID column, rather than being SSN, is now 5-digits long, and has the form of
YYNNN, where YY is the year of hire, and NNN is an integer consecutively
allocated from 000 up within a year. Now also consider the hypothetical hash
function used back in Example 1, where the hashing address is generated
based on the first two characters of the selected hash field. In this case, the
employee ID would not be a good candidate since all the employees hired by
the company in the same year will have the same hashing address.
The number of all possible hashing addresses is too small. For example, it is
not feasible to map millions of U.S. taxpayers’ records to the 50 hashing
addresses representing the 50 states.
Advanced Topics Guide
Hashing theory
Over-allocating disk space
The second collision avoidance technique, which is generally only effective when the
hash key is fairly unique, involves allocating extra space. It is easier to find a hash
function that avoids collisions if we have only a few records to distribute among many
slots than if we have more records than slots. For example, if there are 1,000 slots
reserved and only 100 locations will be used to store records, the result might be very
good using the hypothetical hash function in Example 1. However, the disadvantage is
that the storage space is wasted (90% of the space is wasted in this case). The question
is how much extra space can be tolerated while still obtaining acceptable hashing
performance. In the next section a technique for measuring the performance in terms
of expected overflow will be shown.
Hashing disk space utilization and performance
When creating a hashed table, the database designer must consider the inter-related
design issues of bucket size and packing density, both of which affect database
Bucket size
Bucket size is the number of records that can be held in a bucket. For instance, since a
bucket of 1,024 bytes can store 10 records each of which is 100 bytes long, the bucket
size is 10. Alternatively, if the record length is doubled from 100 bytes to 200 bytes,
the bucket size is decreased from 10 to 5. As bucket size increases, the probability of
overflow decreases but the time taken to search for a record in the bucket may
increase. However, the cost of accessing a record once the bucket is in main memory
is very small compared with the time required to read a bucket in secondary storage.
Packing density
One characteristic of hashing is that the number of I/Os required to perform an
insertion or retrieval tends to increase as the file fills. This seems intuitive since when
most of the buckets are full, it is more likely that the hashing function will generate
the address of a bucket which is already full. Overallocation of the file is an obvious
way to lessen this problem, however it is not desirable to allocate space for a lot of
empty buckets. A measurement called packing density is the ratio of the number of
records (R) in a file to the number of available slots (N): We know that N=B*S where
PackingDensity = -----------------------------------------NumberOfSlots
B is the number of home buckets (which equals the number of addresses that can be
Advanced Topics Guide
Hashed Indexes
generated by the hashing function), and S is the bucket size (the maximum number of
rows which may be stored in each bucket).
PD = ------------B×S
Therefore, since, we also know that we can calculate the number of buckets (B)
required to store a particular hash table, given the number of rows R, and the bucket
size S (which is a function of the page size and the row length) after choosing a
packing density PD by solving the previous formula for B, which gives
 R
---- 
S 
B =  --------- 
 PD 
This concept is important because in SQLBase bucket size is a function of the page
size (which is fixed for a specific implementation platform), the row length, and the
PCTFREE specification on the CREATE TABLE statement. Most of these
parameters cannot be easily changed for hashing performance reasons, but are
determined by other factors. Your main parameter for tuning the performance of a
hash table in SQLBase is by altering the number of buckets, B, in the table’s storage
space. The number of home buckets (B) is the most easily adjusted, and therefore the
most important, parameter which affects packing density. By increasing the number
of home buckets, packing density drops and hashing performance improves due to a
decrease in the number of overflows.
Packing density itself is one of the most important parameters affecting hashing
performance. As a general rule of thumb, collisions are unacceptably frequent if the
packing density exceeds 70%. This rule can be greatly impacted by the bucket size,
which is determined by row and page size, however. This is why you should
understand how to determine a packing density figure that is appropriate for your
particular table.
Hashing performance analysis
Given the choices for bucket size (S) and packing density (R/N), we can estimate the
number of overflows. We must assume that the hash function is uniformly distributed,
which implies two things. One is that the probabilities associated with a record being
inserted into a home bucket are independent. The other is that the average probability
that a record is inserted into one of B home buckets is -B1--- .
Therefore the probability
that a certain set of r records will be inserted into a given
1-  r
home bucket b is --.
that b is one specific occurrence of all home buckets, B.
Also, r represents an arbitrary subset of all records R, which means that r may contain
any number of records from 1 to B.
Advanced Topics Guide
Hashing theory
Next we wish to calculate the probability of the remaining (R-r) records being
inserted into different home buckets other than b. Let 1 – -1--- be the probability that a
record will not be inserted into a particular bucket (called b). Hence, 1 – -1--- 
the probability that (R-r) records are not inserted in bucket b.
We have now introduced the probability that r of R records will be inserted in a home
bucket (b) as well as the probability that the remaining (R-r) records are not inserted
in bucket b. Therefore, the probability of r records being inserted into bucket b and
the remaining (R-r) records not being inserted into bucket b is the product of  -1---  r
 B
and  1 – --1-  R – r (or  --1-  r ×  1 – --1-  R – r  ).
 B
However, there are different combinations that R records could be taken r at a time.
The number of different ways is:
---------------------------r! × ( R – r )!
Thus the probability of a given bucket b having exactly r records (where r ≤ R ) is
1 r
1 R– r .
p ( r ) = ----------------------------- ×  ---- × 1 – ---- 
r! × ( R – r )!
 B
Given p(r), the probability of exactly j overflow records in a specific bucket with
bucket size S is p(S+j). Therefore the expected number of overflows for any given
bucket is :
p ( S + 1 ) + 2 p ( S + 2 ) + 3p ( S + 3 ) + ... + ( R – S )p ( R ) =
∑ j × p(S + j)
Using this formula, we see that the expected number of overflows for all B home
buckets is:
j × p (S + j )
We can express the above formula as a percentage of the records in the file:
∑ j × p ( S + j)
j --=-----1----------------------------Over flowRatio = -------------R
Since we assume a uniform distribution for any given packing density
---R------- ), any bias toward certain addresses in the hash function is likely to
( PD = --B×S
increase the rate of overflow. Be aware that a figure produced with this model is a
minimum percentage. Given the formulas above, we can derive an expected overflow
against packing density for a particular bucket size. The following table shows the
expected percentage of overflow for different combinations of bucket size and
packing density.
Advanced Topics Guide
Hashed Indexes
The following example illustrates the use of this table in the design of a hash table.
Packing Density (%)
Example 3: A company’s employee information is stored using the hashing
technique. Each employee record is 75 bytes. Each disk block has a
usable total of 938 bytes of space. The DBA therefore calculates the
bucket size as 12 (938/75 = 12). The DBA decides on a packing
density of 70%. He also projects that the system must be capable of
holding up to 50,000 employees’ records. The file would therefore
require (50,000/12)/0.70 = 5953 home buckets. Also, with a packing density of 70% and a bucket size of 12, table 2 predicts that at
least 2.13% of the records would be expected to go to overflow
buckets. In other words, there are 50,000*2.13%=1,065 overflow
records requiring 1,065/12 =89 overflow buckets. Therefore every
5,953/89 = 67 home buckets, on average, would require an overflow bucket.
10-10 Advanced Topics Guide
SQLBase hashing implementation
SQLBase hashing implementation
In this section, the implementation of the hash storage technique in SQLBase is
discussed. We will cover how the location of table rows differs for hashed tables as
contrasted to non-hashed tables. We will also describe how to set an appropriate
packing density in SQLBase. In particular, the desired packing density is an input to
the CREATE INDEX command in the form of a specification of the number of hash
buckets to be allocated.
Location of table rows
In terms of the effect on the subject table’s allocation method for each of the two
techniques (hashed or non-hashed tables), rows are located as follows:
Any table which does not have a clustered hashed index placed on it (a nonhashed table) allocates space for row insert operations in a manner which is
basically chronological. The first rows stored in the table are stored on the
first data page, and so on until the most recently stored rows which are found
on the last data page (this is not exactly true because of measures taken to
avoid lock contention between different transactions). These data pages are
allocated in groups of pages called extents, the size of which is determined by
the EXTENSION environment variable setting.
All of the data pages for a hashed table are pre-allocated at the time the
CREATE CLUSTERED HASHED INDEX statement is executed. Rows are
located within these data pages through the use of a hash function, which
maps the symbolic key value of the index to a particular offset within this
already fixed number of pages. As will be seen, when these pre-allocated
pages fill up, no further data pages are allocated. Rather, any rows which do
not fit on the page which is targeted by the hash function are stored on
overflow pages, which will eventually degrade performance.
Hash function
SQLBase contains a pre-programmed hash function which will provide excellent
results for the vast majority of symbolic keys. This function is internal to SQLBase
and cannot be changed by the database designer in any way. For efficiency, SQLBase
implements much of the hash function using bit operations, such as exclusive-or
(XOR), and bitwise complement (1’s complement). The following procedure is
followed by the SQLBase hash function when transforming a symbolic key into a
hash address.
Advanced Topics Guide
Hashed Indexes
How SQLBase hashes a key to a database page
The hashing function is implemented as a 3-step key-to-address mapping:
Transform the symbolic key into a single binary fullword. This allows for
standardization of the remainder of the process. The symbolic key may be one or
more numeric or character types of any length, where each byte can be
represented as an ASCII code. SQLBase performs the following 3 sub-steps to
convert the symbolic key into a standard fullword (4 bytes). This fullword is then
treated as an unsigned binary integer for the remaining steps.
1A. Divide the key into 4-byte units. This accomplishes the standardization of the
1B. Perform a XOR operation on all of the 4-byte units, XORing them all together
and giving a 4-byte result. This ensures that all of the information contained
within the symbolic key is used to compute the hash key. No parts of the key are
omitted from the process.
1C. The result of 1B is a binary fullword. SQLBase then performs a bitwise
complement on it. This operation, also called a one’s complement, simply “flips”
the value of each bit (ones become zeros and vice versa). This is done to eliminate
any bias that may have survived to this point from the symbolic key due to its
coding format, particularly ASCII characters (which are a fairly small range of
ASCII), or SQLBase internal format numeric fields (which are stored in base
Example 4: In an employee table, an eight-character string YYMMCCCC is
used to represent a hash key. The YY (or MM) denotes the year (or
month) when an employee was hired (an example is “9305”, which
means that an employee was hired in May, 1993), and CCCC denotes a name code for the employee (such as where “EJSM” is a
code for E. J. Smith). For this example we will use a symbolic key
value of “9305EJSM”. The ASCII representation for 9305EJSM is
57, 51, 48, 53, 69, 74, 83, and 77.
Key1 = 57,51,48,53
Key2 = 69,74,83,77
Step 1A. The key is divided into four-byte units.
Step 1B. Key1 is XORed with key2.
57,51,48,53 XOR 69,74,83,77 =
0011,1001,0011,0011,0011,0000,0011,0101 XOR
0100,0101,0100,1010,0101,0011,0100,1101 =
(124,121,99,120 in ASCII)
Step 1C. Now perform a bitwise complement on the result of 1B.
10-12 Advanced Topics Guide
SQLBase hashing implementation
~(0111,1100,0111,1001,0110,0011,0111,1000) =
(131,134,156,135 in ASCII)
2,206,637,191 when treated as a fullword binary integer and
converted to decimal.
Transform the binary fullword into an intermediate integer that has the same range
as the number of page addresses in the table’s space allocation. In this step, the
method used is based on performing integer division, then finding the remainder,
through the use of the MOD operation.
This integer is derived by performing a MOD of the output of step one with a
number A, where A is the number of hash buckets that are allocated. SQLBase
always sets A to a prime number to avoid the possibility of divisors producing a
zero result from the MOD operation.
Example 5: To continue from example 4, assume that 1,753 hash buckets have
been allocated for the employee table. This means that allowable
values for the results of step 2 would lie in the range of 0-1752. Dividing 2,206,637,191 by 1753 gives a result of 1,258,777 with a remainder of 1,110. Therefore, the result of step 2 is 1,110.
Transform the output of step two into the physical page address. This is done by
adding a starting page number, which yields the final hash address of the page
containing the row.
Example 6: Suppose that the first page number used for data storage of the employee table rows is page 60. This means that 60 must be added to
the target bucket obtained in example 5 in order to get the physical
target page number of this row. Since 1,110 + 60 = 1,170, the row
will be stored on page 1,170. Similarly, if a SELECT ... WHERE
keyname = “9305EJSM” was performed against the employee table,
SQLBase would look for the row on page 1,170.
Specifying the packing density
Packing density is the single most important design parameter that you can use to tune
the performance of a hashed table. This is accomplished through careful calculation
of the figure used in either the SIZE “b” BUCKETS or the SIZE “n” ROWS clauses
of the CREATE CLUSTERED HASHED INDEX statement (note that “n” is the
number of rows used by SQLBase to calculate the required hash bucket allocation,
whereas “b” is the number of hash buckets that should be allocated specified directly).
Since SQLBase provides a pre-programmed hash function, and the hash bucket size is
fixed with page and row size, this estimate of the CREATE INDEX statement is the
only major influence that a database designer can exercise over the results of the
hashing process.
Advanced Topics Guide
Hashed Indexes
When determining what number to place in this DDL statement, an optimal choice of
packing density becomes critical in order to reduce the number of overflows while
still utilizing disk space efficiently. While the 70% rule of thumb figure is often used,
we encourage you to look in the expected overflow table explained earlier to find the
packing density that is most appropriate for the bucket size that will be in effect for
each of your hashed tables. By looking at this table, you can decide what packing
density is appropriate in order to obtain an expected low number of overflows. This
packing density can then be used as input to the calculation of the number of buckets
estimate for the CREATE INDEX statement. Remember that it is when the CREATE
INDEX statement (not CREATE TABLE) is executed that a fixed number of buckets
will be allocated for storing the hashed table. Also, SQLBase will round up the
number you specify in the BUCKETS form, or the number it calculates in the ROWS
form, of the CREATE INDEX statement to the next higher prime number.
Simple hash bucket calculation
Since in this manual we assume that you wish to control the options available in
SQLBase to achieve your best possible performance, we only consider calculation of
statement. Use of this form of the statement is encouraged in order to target a
particular packing density that is desired, and simplify the specification of this figure
to SQLBase. The alternative form of SIZE “n” ROWS is more appropriate for quick
implementation of hash tables that are trivial in size and do not have stringent
performance requirements. Calculating the row estimate for a desired packing density
is much more difficult and involved than the bucket calculation, and in the end
achieves the same result. For this reason, row estimate calculation will not be covered
or explained in detail. If you wish to use the row estimate option, we encourage you
to make sure your estimate is sufficiently high to avoid the poor performance that
could result from realizing a large number of hash collisions in a table.
We will now describe how the number of buckets for the CREATE INDEX statement
 R
---- 
is calculated given the packing density. As you may recall, the formula
 S 
B = --------- 
 PD 
allows for the direct calculation of the number of buckets, B, given a particular
packing density (PD), number of rows (R), and bucket size expressed as the number
of rows which may occupy a bucket (S).
The S parameter used in the previous formula is the same as the number of rows per
page which is calculated as an intermediate result when estimating the size of a
database table, as explained in the next chapter of database size estimation. To
summarize that formula, the rows per page of a table are calculated as follows (note
that the ΣColLen gth s is the sum of the length of all columns in the table, for
10-14 Advanced Topics Guide
SQLBase hashing implementation
information on how to find the length of each column, refer to Chapter 11, Estimation
Database Size):
PageSize – 86
S = -----------------------------------------------------------------------------------------------------------------24 + ( 2 × NumCols ) + ΣColLengths × 100
-----------------------------------------------------------------------------------------------------------100 – PCTFREE
Example 7: Suppose a table has 7 columns, an average data width of 66 bytes for
all columns, and is projected to contain 60,000 rows. The database
is running on the SQLBase NLM server, which has a page size of
1024K. The PCTFREE parameter of the CREATE TABLE statement has been allowed to default to 10%. The designer has decided
on a packing density of 70%. Bucket size is
1024 – 86
-------------------------------------------------------------------( 24 + ( 2 × 7 ) + 66 ) × 100
-------------------------------------------------------------100 – 10
= 8.
From the expected overflow table we see that the corresponding expected overflow is 3.94%. Using the formula given above, we can
compute the number of buckets required in the CREATE INDEX
statement as (60,000/8)/70% = 10,715 buckets. In other words,
SQLBase will allocate 10,715 pages for storing hashed data. Alternatively, the designer could choose to have an 80% packing density,
increasing the expected overflow to 6.65%. The number of buckets
in the CREATE INDEX statement would then be (60,000/8)/80%
= 9,375 pages would be required. Note that the use of a higher packing density will save space but that the overflow rate will also increase, in this case from 3.94% to 6.65% while the space saved is
10,715 - 9,375 = 1,340 pages or approximately 1.3 megabytes.
Advanced Topics Guide
Advanced Topics Guide
Chapter 11
Estimating Database Size
In this chapter we:
Explain why estimating the size of a database can be an important task for the
Describe the use of existing spreadsheets that simplify the database size
estimation process.
Cover the sizing formulas for the following database objects in detail:
Non-hashed tables
Hashed tables
B+Tree indexes
SQLBase catalog tables
SQLBase overhead
Perform the complete manual size calculations on an example database as an
Advanced Topics Guide
Estimating Database Size
The purpose of this chapter is to enable a DBA to estimate the total size of a complete
database. The raw data needed to perform this task are the detailed database design,
including tables, columns, indexes, and views, as well as row occurrence estimates
for each table in the database.
An accurate projection of a database’s ultimate size is more critical than ever before,
since gigabyte database systems are becoming increasingly common in today’s
market. A database designer can base his estimates of the growth of the database on
observable and measurable business metrics such as the number of new order
transactions processed daily. This information may be extrapolated to row count
estimates for the various database tables involved in each transaction. Following this,
the required disk capacity can be pre-determined by using the formulas provided in
this chapter. The results can then be used to determine appropriate hardware
configurations, both for the initial implementation of the application as well as for
periodic upgrades in capacity.
The starting point for this estimation is to calculate the total size for every table. To
do so, we need to (1) calculate each column width in a table; and (2) calculate the
table size based on the information derived from(1). Table size is calculated
differently for tables which are the subject of a clustered hashed index, due to the
change this has on the table’s space allocation technique:
If the index is created by specifying the option of CLUSTERED HASHED
in the CREATE INDEX statement, we estimate the size of the hashed file as
the table size. This hashed file is allocated at full size immediately when the
CREATE INDEX statement is executed.
Otherwise, we need to estimate the non-hashed table size. This is the size that
we estimate the table will grow to when fully populated with data. Data pages
for the table will be allocated as they are needed over time, as opposed to all
at once as is the case with hashed tables.
If the table is associated with one or more BTree indexes (indexes which do not
specify the option of CLUSTERED HASHED), then the calculation of the table size
should also consider the space required for the indexes. The total space required for
the table is then the sum of the table itself and all associated BTree indexes.
The database size can then be obtained by summing up all the table sizes, the
overhead of all views defined against these tables, and the overhead of the system
catalog (or data dictionary) tables in the database.
Advanced Topics Guide
Formulas for estimating database size
Spreadsheet usage
A number of spreadsheets are available to help the DBA in estimating table and index
sizes. These spreadsheets are available for download from the Centura Web site.
Users of the hashed table spreadsheet, for example, can specify the percent free
parameter (PCTFREE), number of rows projected to be held in the table, and packing
density. Other inputs include column names, data types, average data length for each
column, and declared data length of each column. The spreadsheet calculates the
number of rows contained in a page, number of row pages required, and number of
rows for the CREATE CLUSTERED HASHED INDEX statement. With the
availability of these spreadsheets, the task of estimating the size of a database is
largely automated. The following example in Figure A illustrates one of the
Three specific spreadsheets are available:
Calculation of space requirements for a standard, non-hashed table.
Calculation of space requirements for a hashed table (one which has a
CLUSTERED HASHED index placed on it). This is the example in Figure A.
Calculation of space requirements for a BTree index.
Note that there is no spreadsheet available for calculating space requirements for a
CLUSTERED HASHED index itself. This is because there is no physical space
allocated uniquely to this type of index, rather the effect of the index is to change the
storage allocation method used for the subject table.
Formulas for estimating database size
The database size can be calculated by summing up all the tables, indexes, views, and
a fixed amount of overhead, as follows:
DatabaseSize = ∑RegularTables + ∑ΗashedTables + ∑Indexes + ∑Views
+ FixedOverhead.
Before the size of any table may be calculated, the width of each column must be
determined. Therefore, we first introduce the formula for calculating the column
width. Then, formulas for computing sizes of various tables, indexes, and views are
Note that a page is fixed at different sizes for the various implementation platforms of
SQLBase. In DOS, Windows (with DOS), Windows NT, OS/2, and NetWare the
page size is 1,024 (or 1K) bytes. For reasons of simplicity, we will use 1,024 bytes as
the page size in the examples of this chapter. You should remember to substitute the
page size that is appropriate for your environment.
Advanced Topics Guide
Estimating Database Size
Figure 11A - An example of a spreadsheet which calculates the estimated size of a table.
Note that this table is the subject of a CLUSTERED HASHED index, therefore one of the
calculations made is for the row count to be used in the CREATE INDEX statement.
Also note that the following formulas make use of the floor and ceiling functions. When denoting
these, x is a floor of x, which truncates the fraction part of x (such as, 3.1=3). Also, x is a
ceiling of x, which rounds up the fraction part of x (such as, 3.1=4).
Calculation of column width
A specific formula used to compute a particular column’s width is dependent on its declared data
type. The width computed is the internal width the column will have when SQLBase converts it
into the internal format used by its data type. These individual column widths must then be summed
in order to determine the size of the data portion of a table row.
The input to these formulas is the average external column width of each column. This is the
number of characters or digits that is typically found in the column, as opposed to the declared, or
maximum, size of the column. All data types are stored as variable length fields internally in
SQLBase, so for the column width calculation to be accurate for the majority of the database rows,
the average width must be determined and used. The determination of the average width may be
Advanced Topics Guide
Formulas for estimating database size
done through either a “best guess” method or through detailed analysis of actual
column contents, whichever is more feasible.
Also, hashed tables will need another calculation done using the maximum width for
each column in the table. This will be needed later in order to accurately calculate the
packing density for the hashed file. Since SQLBase reserves space for hashed tables
based on the row estimate in the CREATE CLUSTERED HASHED INDEX
statement, multiplied by the maximum possible row width (which is the sum of the
maximum widths of all non-LONG VARCHAR columns), the calculation of this
maximum row width allows the DBA the highest level of control over the this space
allocation process. This calculation does not need to be performed for non-hashed
Data type
Column width
Number of characters in the column (maximum 254 bytes). For example, a
character type data “Atlanta” is 7-bytes long.
(Note: Physically, SQLBase stores both CHAR and VARCHAR columns the
same way. However, you should continue to use both data type to explicitly
document the logical data type of the attribute. For example, CHAR(8)
implies the attribute always contains eight characters; whereas, VARCHAR(8)
means the attribute may contain up to eight characters.)
(NumberOfDigits+2)/2+1 bytes. For instance, for a 10 digit integer, the
length is (10+2)/2+1=7 bytes.
5 bytes.
12 bytes.
Long varchar
12 bytes consisting of the first physical page number, last physical page number, and column size are stored in a base row page. These 12 bytes are considered as the column width stored in a base row page. The actual data resides in
long varchar pages. SQLBase stores only one column on each long varchar
page. For example, if a column is declared as LONGVAR but its length is only
one byte, an entire long varchar page is allocated to store this one-byte long
varchar data. A general formula for the number of long varchar pages required
for the data is (for each long varchar column):
Nbr of Long Pages per Long Column = size of long varchar in bytes/(102461).
The number of long varchar pages is included in the calculation of a table size.
(Note: SQLBase stores zero length strings defined as either CHAR or VARCHAR as NULL (i.e., requires no physical storage); on the other hand, columns defined as LONG VARCHAR require a minimum of one LONG
VARCHAR page (i.e., 1024 or 2048 bytes depending on the page size.)
Advanced Topics Guide
Estimating Database Size
We define DataLength as the sum of all column widths in a row, which is used for the
table size calculation.
DataLength=∑All Column Widths.
Table size calculations
Based on the derived DataLength, we can compute the estimated regular (or nonhashed) table size or the estimated hashed table size, whichever is appropriate for the
table in question.
Calculation of regular table size
To compute the size for a regular table, we first obtain the DataLength derived from
the previous discussion. We must also have an estimate of the number of rows, called
NbrOfRows below, which will be stored in the table. From these two inputs, we can
calculate the following size estimates.
Row Length
Includes the row overhead, such as row header, and slot table entry. Row
Length = 18 + (2 x number of columns) + DataLength.
Row Length with Slack
Considers the PCTFREE parameter specified in the CREATE TABLE
statement (which defaults to 10%). Row Length with Slack = Row
Length x 100 /(100 - PCTFREE)
Usable Row Page Size
In SQLBase, page overhead is 86 bytes.
Usable Row Page Size = 1024 - 86 = 938 bytes.
Rows per Page
(Usable Row Page Size / Row Length with Slack).
If this formula evaluates to less than 1, then make it 1.
Nbr Row Pages
(NbrOfRows / Rows per Page), where NbrOfRows is the projected
number of rows in the table.
Nbr Long Pages
NbrOfRows x Νbr Long Pages per Long Column.
Total Data Pages
Nbr Row Pages + Nbr Long Pages
Calculation of hashed table size
Computing the size of a hashed table is similar to that for a regular table. However,
we need to take into consideration the desired packing density when estimating a
hashed table size. Therefore a new result parameter must be calculated as follows:
Nbr Hashed Table Pages = Nbr Row Pages / packing density.
Advanced Topics Guide
Formulas for estimating database size
This calculation appears at the end of table 4, which follows.
Row Length
18 + 6 + (2 x number of columns) + DataLength.
The “6” in this equation accounts for storage of the processed hash key with the
Row Length with Slack
Row Length x 100 / (100 - PCTFREE)
Usable Row Page Size
1024 - 86 = 938 bytes.
Rows per Page
(Usable Row Page Size / Row Length with Slack).
Nbr Row Pages
(NbrOfRows / Rows per Page).
Nbr Long Pages
NbrOfRows x Νbr Long Pages per Long Column.
Nbr Hashed Table
(See packing density discussion above for detailed explanation)
Total Data Pages
Nbr Hashed Table Pages + Nbr Long Pages
Nbr Hashed Table Pages = Nbr Row Pages / packing density.
BTree Index size calculations
When a table is associated with a BTree index through the CREATE INDEX
command, we need to calculate the BTree index size. To perform this computation,
we first obtain the length of the index’s symbolic key by performing the column width
calculation for each column named as a subject of the index. We must also have an
estimate of the number of rows, called NbrOfRows below, which will be stored in the
subject table. From these two inputs, we can calculate the following size estimate for
the lowest level of an index (leaf pages of the index structure) as follows:
Key Length
Key Length = ∑ average data lengths of columns in the key.
Index Entry Length
9 + number of columns in index + Key Length.
Usable Index Page Size
(1024 - 74) x (100 - index PCTFREE) / 100 (Note that the default value
for PCTFREE is 10%).
Index Entries per Page
(Usable Index Page Size / Index Entry Length).
Nbr Index Pages
Nbr of Rows / Index Entries per Page .
Advanced Topics Guide
Estimating Database Size
Estimation of overhead for views
For each view there is fixed overhead and a variable overhead depending on the
complexity of the view. Note that when a view references another view, the variable
overhead of the referenced view is added to the view being defined. For example, if a
view refers to two simple views whose variable overhead is 5 pages each, the
complex view will have a body which is 5+5=10 pages plus its own variable
overhead. In other words, the variable overhead of each referenced view is included
in the variable overhead of the view that references it.
The size of a view can grow rapidly if it references other views which in turn
reference yet other views.
Fixed Overhead
12 x 1024 bytes (12 pages)
Variable Overhead
((Nbr of Tables x 150) + (Nbr of Columns x 170)) bytes
Variable Overhead of
other views
The sum of the variable overhead of all views referenced by this
∑ Variable Overhead for all views
Total View Overhead
in Pages
(Fixed Overhead + Variable Overhead + Variable Overhead of other
views) / 1024 pages.
Fixed amount of overhead
This calculation accounts for the overhead of the system catalog (or data dictionary)
tables in the database. SQLBase maintains the skeleton system catalog tables in a file
called START.DBS. SQLBase uses the file as a “template” for a database file. Each
new database is created originally from a copy of START.DBS (the CREATE
DATABASE command and the API call sqlcre automatically copy START.DBS to the
database named in the command). The size of START.DBS therefore has to be
included in the fixed overhead estimate for a database. This size may change from
one release of SQLBase to another, and should be obtained from the current version
when performing the calculations. For our examples, we are using the size of
START.DBS from SQLBase version 5.1.4.
Advanced Topics Guide
Formulas for estimating database size
There is a fixed amount of overhead for various physical objects. If the object is either
a regular or a hashed table, the overhead is 12 pages. For non-hash indexes, the
overhead is 2 pages. Therefore, the formula for the total amount of system overhead
for a database is:
Total Fixed Overhead Pages = (Nbr Tables x 12) + (Nbr non-hash
Indexes x 2) + (Size of START.DBS / 1024).
In actuality, there is also some variable overhead associated with each type of object
which depends on the current size of the object. As the size of the object increases, so
will this overhead rise slightly. For example, the number of bitmap pages which
SQLBase uses to manage page allocation for data rows will increase with the number
of rows, however this increase is very small, at approximately 1K for every 8MB of
data. Because of the tiny incremental nature of this overhead, formulas are not
presented here to accommodate their inclusion in the overall estimation.
Advanced Topics Guide
Estimating Database Size
Example of size estimation
We present a simple database to explain the estimation of its size using the derived
formulas. Assume that a packing density of 70% is desired for the CUSTOMER
table, which is the subject of a hashed index.
Syntax Key
total bytes
rec len
key length
occ urances
Pct free
# pages
prim ary key
foreign key
colum n 1
colum n 2
U nique N o n-uniq ue H as h ed
Figure 11B - Diagram of database following size calculations. All spaces
concerning size information, with the exception of row occurrence estimates, would
only be filled in following completion of the size estimation exercise. Note that the
only data derived directly from the size estimates is the number of pages for each
table. The “rec len” and “total bytes” fields reflect the external width of the
11-10 Advanced Topics Guide
Example of size estimation
The SQLBase Data Definition Language (DDL) statements used to create this
database, and associated views, follows:
SIZE 47628 ROWS;
After an analysis of the data that will be loaded to the database was performed, the
following data on average column widths was derived:
Advanced Topics Guide
Estimating Database Size
The estimation of the database size is performed as follows:
Data Length=∑Column Widths of Customer_ID, Customer_Name,
Customer_Addr, Customer_Rating.
Data Length=5+10+30+5=50
Row Length=18+6+(2x4)+50=82
Row Length with Slack=(82x100)/(100-15)=97
Rows per Page=(1024-86)/97=9
Nbr Row Pages=50,000/9=5,556
Nbr Hashed Table Pages=5,556/.70=7,938.
(Note that the .70 figure is desired packing density)
Since there are no LONG VARCHAR columns in this table, the total data
pages equals the number of hashed table pages.
Data Length=∑Column Widths of Customer_ID, Contact_Name,
Contact_Phone, Contact_Text.
Data Length=5+15+(((10+2)/2)+1)+12=39
Row Length=18+(2x4)+39=65
Row Length with Slack=(65x100)/(100-25)=87
Rows per Page=(1024-86)/87=10
Nbr Row Pages=175,000/10=17,500
Nbr Long Pages=175,000x1=175,000
Total Data Pages=17,500+175,000=192,500
11-12 Advanced Topics Guide
Example of size estimation
The following is for the BTree index calculation for XPK_CONTACT.
Key Length=5+15=20
Index Entry Length=9+2+20=31
Usable Index Page Size=(1024-74)x(100-10)/100=855
Index Entries per Page=855/31=27
Nbr Index Pages=175,000/27=6482
The first calculation is for the BAD_CUSTOMER view:
fixed overhead=12x1024=12288 bytes
variable overhead= (1x150)+(2x170)=490 bytes
variable overhead of other views = 0
total overhead of the view=(12288+490+0)/1024=13 pages
The next calculation is for the BAD_CONTACTS view:
fixed overhead=12x1024=12288 bytes
variable overhead= (2x150)+(3x170)=810 bytes
variable overhead of other views = 490
total overhead of the view=(12288+810+490)/1024=14 pages
Fixed amount of overhead:
(2x12)+(1x2)+(602,112/1024)=614 pages
Thus, the estimated size for the database is:
7,938+192,500+6,482+13+14+614=207,561 pages.
Since each page is 1024 bytes, or 1K, and one megabyte is 1024K, the size in
megabytes is the number of pages divided by 1024:
207,561 / 1024 = 203 megabytes.
Advanced Topics Guide
Advanced Topics Guide
Chapter 12
Result Sets
In this chapter we:
Introduce the concept of the result set, which is the intermediate storage area
used by SQLBase when processing a SELECT statement which retrieves
multiple rows from one or more database tables.
Cover the following detailed characteristics of result sets:
Life cycle
Evaluate the implications that SQLBase’s result sets have for program
processing requirements, especially in the areas of locking and concurrency.
Advanced Topics Guide
Result Sets
A result set is a temporary index, possibly associated with a temporary table for
certain complex queries, which contains a set of data that represents the result of a
SELECT statement. Since most SELECT statement executions cause a number of
rows to be returned from SQLBase, the result set typically consists of many rows.
The contents of the result set’s leaf pages (the bottom level in the B+ index tree
structure) are pointers (ROWIDs) to the rows of the table (or tables) involved in the
SELECT. Alternatively, for some complex queries these ROWIDs point to the rows of
a temporary table containing the data values needed to satisfy the SELECT statement.
The result set is stored on disk in a temporary file, and one or more rows are
transmitted from it to the client workstation when requested via the FETCH
The utilization of result sets is an important feature of SQLBase. Through the use of
result sets, SQLBase decreases response times for client workstation users while
increasing their flexibility to position themselves within the result set when
performing browse operations. SQLBase also offers a number of options relating to
the processing of result sets which allow the system designer to balance the flexibility
of result set processing with the concurrency requirements of his application.
Characteristics of result sets
The structure of result sets is standardized in a manner to allow their efficient use by
the SQLBase server. However, the content of a result set as it exists on the server is
determined in part by the nature of the SELECT statement which created it. The
timing of the creation of result sets depends in part on the manner in which rows are
fetched from it by the program which issued the SELECT statement originally.
Physical structure of result sets
The structure of result sets as they are stored in the server is that of a B+Tree index
similar to any other SQLBase index. However, the result set is not a permanent object
as user-defined indexes are and does not possess a name, nor is it owned by any
database. It is owned by the transaction that created it, and is associated with the
cursor used to process the SELECT statement whose results it represents. When this
cursor is disconnected, or used to prepare another SQL statement, the result set is
dropped from the system.
The subject of the unique index structure in a result set is the ROWID column
contained within each leaf entry. To summarize chapter 8, Database Pages, the
following ordered list of elements makes up the ROWID field, which serves to
uniquely identify every row in an SQLBase database:
Physical page number that the row resides on.
Advanced Topics Guide
Characteristics of result sets
Slot number within the page that the row occupies.
Row serial number, which is assigned at the time of insertion and is unique
among all rows. This number is never updated during the life of the row, and
is never reused for any other row following the deletion of this row.
Update serial number of the row. This number is changed each time the row
is updated.
Note that the first three elements will always uniquely identify a single row, while the
addition of the fourth element serves to further identify that unique row for a single
state, or change level. When the row changes to a new state, through being the subject
of an UPDATE command which modifies one or more columns, then the fourth
element of the ROWID changes while the first three remain the same. Also, only the
first two elements are needed by SQLBase to actually retrieve the row from storage.
The third element serves to verify that the same row still occupies this position in the
database, and that the slot has not been reused by some other row following the
deletion of this row.
Logical contents of result sets
Result sets are composed of ROWIDs which point to either base table rows, or rows
in a temporary table, depending on the query that was executed. Result set columns
contain base table ROWIDs when the query references data columns as they appear in
the actual table rows, or when a calculation or formula is to be applied to a single row
column value. These ROWIDs are then used at fetch time to access the base table row
where the data is located. Result set ROWIDs point to a temporary table when the
column to be returned is a value calculated from more than one input, or an aggregate
function, which may use one or more rows of the base table as input.
Advanced Topics Guide
Result Sets
Select Ac ctnum , Acctnam e from C ustome r;
R esu lt set leaf en tries
RO W ID of row 1
RO W ID of row 2
RO W ID of row 3
RO W ID of row n
Figure 12A - Result set from a single table query with no derived data. The only entry
in the leaf pages of the index tree is the ROWID of each of the Customer table’s rows
which qualifies for inclusion (in this case, all rows).
In Figure A the result set of a simple one-table query is shown. For SELECT
statements such as this, the result set consists of the ROWIDs of the selected rows in
the Customer table. When the client performs a fetch against the result set, the row or
rows returned are identified by the ROWID at the cursor position. As the data rows
are accessed, the columns requested by the query are retrieved and formatted for
return to the client.
Advanced Topics Guide
Characteristics of result sets
Select C.A cctnum, C .Acctna me, I.Invoice_am t
from C ustom er C , Invoice I
w here C .Acc tnum = I.Acctnu m;
R esu lt se t leaf entries
Custom er
ROW ID for row 1 RO W ID for row 1
ROW ID for row 2 RO W ID for row 2
ROW ID for row 3 ROW ID for row 3
ROW ID for row n RO W ID for row n
Figure 12B - Result set from a two table join. ROWIDs identify the qualifying rows of
each table, and allow SQLBase to retrieve the appropriate columns from the selected
rows of the base tables at FETCH time.
In figure B, the result set of a simple two-table join is shown. Since all of the data
columns requested by the query are present in the rows of the two tables, the only
columns in the result set are the ROWID pointers for the joined rows. Each ROWID
pair in the result set satisfies the join condition, which is that the account number in
the Customer row matches the account number in the Invoice row. When the actual
rows are fetched by the client, SQLBase will use the ROWIDs to retrieve the actual
data values requested by the query from both tables involved in the join. Note that this
query will result in one row for each invoice a customer has on file, in other words
Advanced Topics Guide
Result Sets
each customer number will be repeated in multiple rows for each invoice that
customer has.
Select C .Acctnum , C.Acctname, SUM(I.Invoice_am t)
from C ustom er C , Invoice I
where C.Acctnum = I.Acctnum
group by C .Acctnum , C.Acctname;
Result set tem porary table
Resu lt set leaf en trie s
C olum n 1
Co lum n 2
Tem po rary table
C.Acctnam e
Colum n 3
S U M of
I.Invoic e_ am t
ROW ID for row 1
VALUE for row 1 VALUE for row 1 VA LUE for row 1
ROW ID for row 2
V ALUE for row 2 VALUE for row 2 VALUE for row 2
ROW ID for row 3
V ALUE for row 3 VALUE for row 3 VALUE for row 3
RO W ID for row n
VALUE for row n VALUE for row n VALUE for row n
Figure 12C - Result set from a two table join aggregate query. Since this query
contains aggregate data in the form of the SUM (I.Invoice_amt) column, a temporary
table is built to hold the result set data. The result set itself points to the rows of this
temporary table. This temporary table is also sometimes called a ‘virtual result set.’
The aggregate query in Figure C has a significant difference from the previous
queries we have examined, in that its result set contains the ROWIDs of a temporary
table. This temporary table contains the actual row values to be sent back to the client
in response to fetch statements. These values must be stored in a temporary table
otherwise the entire aggregate set of data would have to be retrieved again in order to
recalculate the result of the function. The query in figure C contrasts with the query in
figure B in that it returns a single row for each customer. This row will contain the
total value of all invoices for that customer, calculated by the SUM function which is
applied to the invoice amount field. SQLBase creates a temporary table for any query
that includes aggregate functions, complex views, or queries containing the
Advanced Topics Guide
Characteristics of result sets
This delineation of queries that require calculations, aggregation, or other complex
processing determines variations in other characteristics of result set processing.
Result sets which contain the ROWIDs of base tables possess a greater flexibility for
processing by SQLBase. These result sets offer some capabilities to client
transactions that the other result sets do not. For the remainder of this chapter, the
differentiation between these two types of result sets will be made through the term
“derived”. When a result set contains the ROWIDs of base tables only, where no
temporary tables are built, all the data needed to satisfy the query can be retrieved
from the base table rows. More complex queries, such as aggregation queries, have
data derived from the processing of the query present in the temporary table, pointed
to by the result set, for subsequent transmission to the client.
To summarize, the two basic types of result sets are:
The non-derived result set, which contains only ROWIDs of data rows in the
base tables.
The derived result set, which contains a ROWID pointer to a temporary table
that contains at least one column of actual data values which represent the
result of some calculation or function in the query. This type of result set is
also called a “virtual table”.
Population of result sets
For most queries, the first part of the result set is built when the SELECT statement is
the subject of an EXECUTE. When there are sufficient rows to fill one message buffer
to the client, these rows are transmitted and the client may proceed. After the client
has performed sufficient FETCH operations to exhaust the contents of the message
buffer, a request is made to the server for another buffer of rows. The server then
continues to build sufficient rows in the result set to fill another message buffer, which
is transmitted to the client workstation. This process continues until all rows that are
eligible for inclusion in the result set have been accessed. Note that this process
assumes a sequential access of result set rows by the client workstation’s program,
starting at row 1 and proceeding through the result set, one row at a time, until all
rows are retrieved.
Of course, the program at the client workstation may not proceed through the result
set in this orderly manner. When the program explicitly repositions in the result set,
SQLBase may be required to complete building the entire result set prior to sending
the requested row back to the client. When this occurs, the user at the client
workstation may perceive a delay if the result set that must be completed is large.
Also, some queries may require the completion of the entire result set prior to
returning any rows. An example of this is a SELECT statement containing the
ORDER BY clause. This clause requires that the result set rows be ordered according
to the contents of one or more columns. If there is an index on the base table which
Advanced Topics Guide
Result Sets
matches the sequence of the ORDER BY clause, then the result set may be built
incrementally, as outlined above. However, if there is no index that can be used to
satisfy the ORDER BY clause, then SQLBase will have to perform a scan of the
entire table, followed by a sort operation, in order to put the result set in the correct
sequence. Only following this sort may the first row of the result set be returned to the
Flushing of result sets
When result sets are no longer needed by the program that created them, they are
flushed from the server. This is analogous to dropping a database object, where all
traces of the object are eliminated from the system and there is no way to bring the
object back. If the result set is large, the temporary file used by the server for its
storage will also be freed. As mentioned previously, the result set table is not a
permanent object nor does it possess a name or ownership by any database. It is
owned by its creating transaction, and is associated with the cursor used to process
the SELECT statement. In the absence of any special considerations taken to preserve
the result set, the result set will be flushed when one of the following events occurs:
When the owning cursor is disconnected, or used to prepare another SQL
When the program re-compiles the SELECT statement.
When the transaction requests a COMMIT or ROLLBACK.
When the transaction changes the current isolation mode setting.
Cursor context preservation
If the programmer has turned cursor context preservation (CCP) on, by setting the
SQLPPCX flag on, then the above rules for the flushing of the result set are changed
somewhat. Specifically, the program’s result sets are maintained across a COMMIT,
allowing the program to perform update operations and commit the results without
losing its current place in the result set. Also, for programs running in the RL
isolation level, and not performing DDL commands, result sets will be kept intact
across the execution of a ROLLBACK command.
This flag allows programs to browse large result sets, perform updates based on the
result sets, commit or rollback those updates as desired, and then continue their
browsing without having to reestablish the result set. This makes these programs very
effective in a multi-user environment, since the commit allows locks to be freed up
which other transactions may be waiting on, while the CCP feature allows the
transaction to avoid the overhead of building the same result set multiple times.
Advanced Topics Guide
Processing result sets
Processing result sets
You need to be aware of the implications that SQLBase’s use of result sets have for
transactions. While the use of result sets can significantly increase the speed of
transaction execution, programs occasionally need to take steps to avoid some
anomalies caused by result set usage. These aspects of result set processing may seem
all the more odd because of the behavioral variations between derived and nonderived result sets, and the fact that many people may be unaware of the basic
processing differences in these types of queries.
Result sets are built at a specific point in time during the execution of a program,
namely when a SELECT statement is executed. However, the data contained within
the result set may be perused by the client workstation user over a more extended
period of time, particularly when browsing a large result set. This raises two questions
relating to concurrency:
What locks are held on the rows that were used by the SELECT to build the
result set? This will affect other transactions’ ability to access these rows.
What happens if some other program updates a table row that was used to
build the result set? This affects the timeliness of the data being returned to
this transaction. Another way to ask this question would be: Does the
information in the result set represent data at the current point in time, or at
the point in time that the SELECT was processed?
Note that some transactions may not be too concerned with these issues. They may
only need a result set which is representative of the data at the time of the SELECT
and do not care if the data has changed since that time. These transactions should
choose a high concurrency isolation mode for execution (probably RL), so that other
transactions may perform updates without interference. The transaction may then
browse its result set without concern for the timeliness of the data.
An example of a program that may not require a high degree of concurrency is an
“orders shipped today” inquiry screen in an order processing system. A user viewing
this screen would not realistically expect to see an order on it that just went out the
door while he is browsing the result set of orders. The probable reason the inquiry
exists is for management to identify trends and potential problems with the order
shipping process and this use does not require that the screen reflect every single
order in real time. The accurate representation of “orders shipped today” would be a
batch report program that was run at the stroke of midnight, thereby guaranteeing that
no more orders could be added to the desired information.
Advanced Topics Guide
Result Sets
Lock duration on result set rows
How long locks are held by result sets is determined by the executing transaction’s
isolation mode. The possible isolation modes are Repeatable Read (RR), Cursor
Stability (CS), Release Locks (RL), and Read Only (RO). While these are discussed
in detail in chapter 13, Concurrency Control, their effects on result set processing are
summarized below.
In RR, S-locks remain on all pages where rows from the result set are located
throughout the life of the transaction, or the next commit or rollback. While
other transactions may read these rows, they will be unable to update them
until the S-locks are released.
In CS, an S-lock is only held on the current page where the most recent row
FETCHed by the client is located. Since the server passes only one row at a
time to the client in this mode, the page with the S-lock will be the page
occupied by the row last passed in the message buffer. The S-lock will move
to the page occupied by the next row in the result set when another FETCH
is performed by the client. This means that a row in a page of the result set
could be updated by another transaction so long as it is not the page
containing the current row of the result set.
In RL, no locks are held on result set rows permanently. During the actual
execution of the SELECT statement, to build the result set, one S-lock is held
on each page as it is accessed, but this lock is dropped as soon as the page is
no longer needed. Once the result set is built, all locks are released. However,
each row in the result set is subject to one more S-lock at the moment the
fetch of that row is performed in order for SQLBase to make sure the row is
not the subject of an uncommitted update from another transaction at that
point in time. This S-lock is immediately released when SQLBase finishes
the fetch operation for that row.
In RO, no locks are ever held. Row images are always made current to the
SELECT execution time through the use of timestamp comparisons and a RO
history file which holds the prior versions of row images.
Concurrency and derived result sets
For queries which create derived result sets, or virtual tables, the remainder of the
concurrency situation is fairly simple. Since the data that makes up the results of the
query is actually stored in the temporary table, any updates made by other
transactions after the result set is built will not be reflected, nor can these updates be
detected. There is no process that can avoid this result, such as the use of
FETCHTHROUGH or committing of the update. Also, committing the update with
cursor context preservation on will not cause the update to be reflected in the result
12-10 Advanced Topics Guide
Processing result sets
set either. The only way to make the results of the update appear in the virtual table is
to re-execute the SELECT statement which built it.
Concurrency and non-derived result sets
On the other hand, if the result set contains no derived columns, a transaction can
detect updates to rows, or deletions of rows, made after the result set was built. For
this type of query the following sections describe the techniques and procedures for
dealing with these possible updates and deletions made by other transactions. Note
that in the case of insertion of new rows by other transactions that would have
appeared if they had been committed prior to the creation of the result set (meaning
that they satisfied the predicates specified in the WHERE clause), these new ROWIDs
will not appear in the result set under any conditions. The only way to include them in
the result set would be to re-execute the SELECT statement in order to re-evaluate the
predicates against the existing set of committed data in the table.
Detecting changes in result set rows
Transactions which execute in the RR isolation mode will not have this problem,
since this mode will not allow any updates to occur to the result set. While this may be
appropriate for some transactions, they should only be executed during low
concurrency time periods, such as the middle of the night, otherwise a large number
of time-outs and deadlocks may occur. For transactions which are processing result
sets in the CS or RL isolation modes, there is a way to keep the data returned from the
result set current.
SQLBase provides this capability automatically when fetching rows from the result
set. Since this is done at fetch time, any rows that have already been fetched, and then
updated by a different transaction at a later point in time (following the fetch), will not
be detected. At the time that the fetch is performed from the result set, SQLBase uses
the ROWID stored in the result set to verify the current state of the row as it is actually
stored in the base table. It does this as part of the following process:
How SQLBase fetches result set rows
Access the ROWID stored in the result set row.
Retrieve the page using the page number component of the ROWID. This page
could either be found in a buffer or be read from external storage.
Check the slot specified by the ROWID slot component to make sure the “used”
bit is set (first bit in slot table entry). If this bit is off then the row has been deleted.
If this has occurred, then SQL return code 00003 (FET NF: Row has not been
found) is returned to the client.
Retrieve the row from the page using the offset specified in the slot table. This
structure is explained in detail in Chapter 8, Database Pages.
Advanced Topics Guide
Result Sets
Compare the row serial number component in the ROWID to the row serial
number actually stored in the slot. If these are different, then the original row has
been deleted and the slot has been reused by another row. If this has occurred, then
SQL return code 00003 (FET NF: Row has not been found) is returned to the
Compare the update serial number component in the ROWID to the row update
serial number actually stored in the row. If these are different, then the row has
been updated since the SELECT statement that built the result set was processed.
If this has occurred, then SQL return code 00002 (FET UPD: Row has been
updated) is returned to the client, indicating the row has been updated. The newly
updated row is returned to the client.
This procedure ensures that transactions processing result sets in high concurrency
isolation modes can still detect modifications to the rows in their result set which
occur after the result set is built.
Maximum timeliness with FETCHTHROUGH
The process just described is only effective if the row fetched by the client is not in
the workstation’s input message buffer. If the row is already in the client’s buffer
when an update takes place, then the transaction will not be able to detect it. In order
to eliminate this effect, the SQLBase FETCHTHROUGH feature can allow a
transaction to execute with maximum concurrency in an isolation mode such as RL,
while still accessing the most recent data available. By setting the
FETCHTHROUGH parameter on (via session setting or API call), the transaction can
ensure that for each FETCH performed, the data will be retrieved directly from the
server, who will perform the automatic recency verification described above. The
alternative to this, with FETCHTHROUGH set off, would be that the row may be
retrieved not from the server at all, but rather from the client’s own message buffer.
The contents of this buffer may have been built some time before, and no longer
contain the most recent information. Because of the performance impact of using
FETCHTHROUGH, however, you should only turn it on when you know that the
result set rows will be updated during the life of the result set.
An example of a transaction that may require this high level of timeliness is a
shipment pick transaction for a high volume warehouse. The way the transaction
might work is that while a stock clerk makes his way down an aisle in the warehouse,
he scans the bin numbers he passes with a bar code reader. The transaction being
executed on his hand held computer has already performed a SELECT on the
SQLBase server (connected via radio frequency modem) and built a list of items and
quantities to be picked in the same order that the clerk will traverse the warehouse. As
the clerk scans a bin, the transaction executes a FETCH of that bin’s row in the
Warehouse table to see how many of the bin’s items should be picked by the clerk for
conveyance to the packing area. Since orders are being placed constantly, the
12-12 Advanced Topics Guide
Processing result sets
transaction executes with FETCHTHROUGH on in order to be aware of the latest
order information that may affect the quantity to pick from this bin. In this way, the
clerk will be performing warehouse picks for orders which were received after he
started his run through the warehouse. This thoroughness on the part of the shipping
process helps to decrease the time required to turn around a customer’s order, which
helps the company achieve several of the strategic business goals that they purchased
the system to accomplish.
The trade-off involved in guaranteeing the most recent data through use of the
FETCHTHROUGH feature is an increase in network message traffic for the
transaction. This is because instead of filling the message buffer with as many rows as
possible to transmit to the client, the server will now place only one row at a time in
the buffer. This will cause the client workstation to experience slower response times
when processing large result sets. For this reason, FETCHTHROUGH should only be
used on those situations where it is critical to have the most recent information, and
there is reason to believe that the information is in a constant state of change. Also,
anytime you know the row has been updated, such as through an UPDATE statement
issued in another cursor of your program, you should set FETCHTHROUGH on to
ensure that the update will be reflected in the result set.
Advanced Topics Guide
Advanced Topics Guide
Chapter 13
Concurrency Control
In this chapter we:
Introduce the topic of concurrency control, which is the way a SQLBase
server makes each client workstation appear as if it were in sole control of the
Discuss the concept of transaction serialization, which dictates the
requirements for inter-leaving transaction executions in a manner which
eliminates interference between transactions and makes each transaction
individually recoverable.
Explain the various types of locks SQLBase uses to control transaction
executions, and the characteristics of these locks.
Delineate the various isolation levels available in SQLBase, which determine
how long certain types of locks are held during a transaction’s execution.
Explain how to select an appropriate isolation level for a program, balancing
the trade-offs of data integrity and maximum performance for all executing
Advanced Topics Guide
Concurrency Control
The topic of concurrency control addresses the problems faced by SQLBase in
supporting a multi-user environment. When users are at their workstations using
some type of client software, which could be a SQLWindows application, Quest,
WinTalk, or some other program, and are connected to a SQLBase database server,
they expect the results of their programs to be exactly the same as if they were
executing against a local copy of the database contained on their own workstation.
Users want to remain unaware that many other users are also connected to the same
database and are using it in a variety of ways. Unfortunately, there are problems
inherent in multi-user database access that must be considered by both SQLBase and
the application developer in order to realize the users’ objective. These problems are
caused by the pitfalls inherent in allowing many users to access the same database
concurrently. The topic of concurrency control addresses the software mechanisms
used by SQLBase to overcome these difficulties.
The unit of work that is the subject of the concurrency problem is the transaction
(which is discussed at length in Chapter 5, Transaction Definition). One of the key
goals of concurrency control is that either all updates of a transaction are made
permanent if its execution is successful, or it otherwise has no effect at all on the
shared database. The other goal of concurrency control is that transactions not be
allowed to interfere with each other. Each transaction should appear to have sole
control over the shared database without concern for other transactions whose
execution is being conducted simultaneously by SQLBase.
Transaction serialization
Transactions are executions of programs against the shared database which are
delineated by an implicit start when the program connects to the database, and are
ended through either a commit or a rollback command. Transactions which complete
successfully signal this event to SQLBase through the COMMIT command, at which
time the changes made by the transaction are made permanent within the database.
Unsuccessful transactions terminate with the ROLLBACK command, which causes
SQLBase to back-out all changes to the database that were previously performed on
behalf of the transaction. If the program continues to perform database calls after the
execution of a COMMIT or ROLLBACK, a new transaction is considered to have
begun. A rollback of the transaction may occur for a variety of reasons besides
program request (as indicated through the program issuing the ROLLBACK
command) including program failure (abort), system failure (OS abend), or
environmental failure (power failure). If any of these events occur, then SQLBase
must ensure that all non-committed transactions are rolled back so that the database
never appears to have been effected by them.
Advanced Topics Guide
Transaction serialization
A transaction may have two possible effects as a result of its execution:
Effects on data through the execution of INSERT, UPDATE, and DELETE
DML statements.
Effects on other transactions, where the other transactions have read data
written by this transaction. These other transactions may use uncommitted
data to perform further database alterations, thereby proliferating the effects
of this transaction’s updates.
The effects of a transaction that were made directly to the database through the use of
DML statements are eliminated through the performance of a rollback by SQLBase
using the record images that are kept in the log files (the details of this operation are
covered in Chapter 14, Logging and Recovery). Eliminating the secondary effects on
other transactions is where the mechanism of database locks comes into play. These
locks are the mechanism used to ensure that transactions occur in a serialized fashion,
thereby allowing their recovery without impacting the integrity of other transactions
which may otherwise have been dependent on data created by an uncommitted
Databas e
Tra nsaction 1
Transac tion 2
UPDATE order
W HERE ordno = 1
SET status = 's';
Correct restored
im age after both
transactions aborted
S ELECT order
WHE RE ordno = 1;
IF status = 's' THEN
UPDA TE cust-acct
WHE RE acctno = order.acctno1
SE T balance = balance + order.ordtotal;
[not allowed]
Figure 13A - Example of transaction dependency without database locking where
Transaction 2 should not be allowed to commit since it depends on data supplied by
the uncommitted Transaction 1. This is not a serialized execution, and is therefore
only recoverable by rolling back Transaction 2 along with Transaction 1, called a
“cascading abort.”
Advanced Topics Guide
Concurrency Control
An example of secondary transaction effects is shown in Figure A, where transaction
2 has read data from transaction 1 prior to transaction 1 committing. This example
illustrates the impact of not having locks, which is that transaction 2 must be aborted
due to the request of transaction 1 to rollback rather than commit. This cascaded
abort of transaction 2 is required to preserve the semantic integrity of both
transactions, since allowing transaction 2 to commit would propagate inaccurate data.
The alternative of not processing the rollback request of transaction 1 would also
allow inaccurate data to remain in the database.
D atabase
Transactio n 1
UPDATE order
W HERE ordno = 1
S ET status = 's';
Transac tion 2
X - L o ck
w rite
SE L EC T ord er
W HE RE o rdn o = 1 ;
re ad
Correct restored im age
after transaction 1
[w ait for X -Loc k releas e]
IF s ta tu s = 's' T HE N
UP D ATE c u st-a cc t
W HE RE a c ctn o = o rd e r.a c ctn o 1
S ET b al an ce =b a la n ce + ord er. ord to t a l;
[u p d a te no t pe rfo rm e d ]
Figure 13B - The same as Figure A, except with locks. Transaction 1 places an
exclusive lock on the row it updates, which causes Transaction 2 to wait until
Transaction 1 completes prior to reading the record. This is a serialized execution,
and both transactions are fully recoverable through rollback techniques.
Figure B shows the same two transactions executing with database locking
performed. Following the update of the ORDER row by transaction 1, SQLBase
places an exclusive lock on the page containing the row. This type of lock can only be
held by one transaction at a time and blocks all other transactions from accessing the
page (more detail on locks follows later in this chapter). When transaction 2 attempts
to read the ORDER row, via a SELECT statement, it is placed into a wait state until
the exclusive lock is released. During this time, transaction 1 requests a rollback,
which is performed. Following the rollback, the exclusive lock is lifted on the
ORDER row, and transaction 2 is allowed to continue. Notice that the row image
returned to transaction 2 is not the one that existed at the time of its request, as in
Advanced Topics Guide
Lock types and scope
Figure A, but rather the restored image which existed prior to modification by
transaction 1. Since the status value is no longer ‘s’, transaction 2 no longer modifies
the CUSTOMER row as in Figure A, but terminates with a COMMIT anyway. These
transactions are now serialized, they no longer execute in an inter-leaved fashion
which allows them to affect each other’s execution.
This example is one of a number of ways in which transactions may interfere with
each other’s correct execution through their interaction within a shared database.
Other common problems that may occur due to lack of concurrency control include:
Lost Updates
Lost updates occur when two or more transactions retrieve the same record
and make successive updates of that same record. If no method of locking is
used, transactions can overwrite recent modifications without the benefit of
seeing the newly updated record.
Uncommitted Dependency Problem
Rollbacks issued by one transaction can also cause another transaction to lose
updated data if no locking protocol is used. If a record is updated by one
transaction but not committed, then subsequently updated by a second
transaction with no commit performed, a rollback by the first transaction will
also rollback the second transaction's updated data.
In order to eliminate these problems, SQLBase ensures that transactions execute in a
strict serialized fashion through the use of database locks. These serialized executions
achieve the same effect as executing each transaction in serial fashion, with respect to
all other transactions, such as would happen if they were dispatched through a FIFO
(First In First Out) queue. Of course, SQLBase does not accomplish this by FIFO,
since only through concurrent executions can the speed that is required for an
effective RDBMS be achieved. When viewed conceptually, however, for any single
transaction the effect of its execution is exactly the same as if it were executed
serially, without other transactions present in the system.
Lock types and scope
SQLBase employs two mutually exclusive types of locks: Shared Locks (S-Locks),
and Exclusive Locks (X-Locks). An alternative terminology that is sometimes used
refers to S-Locks as read locks and to X-Locks as write locks. The one other type of
lock which SQLBase uses on a temporary basis is the Update Lock (U-lock), which is
sometimes called incremental lock because of its transient nature. These locks are
placed implicitly (that is, without being specifically requested by the transaction) by
SQLBase in response to the various DML commands issued by a transaction, with
one exception. This exception is the ‘FOR UPDATE OF’ subclause of the SELECT
Advanced Topics Guide
Concurrency Control
command, which causes an explicit U-lock to be placed in anticipation of an update
occurring later within the transaction.
With an S-lock, more than one simultaneous transaction can lock the record;
therefore, more than one transaction can access the record. S-locks are applied when
a transaction reads data via a SELECT statement. When data is held by an S-lock,
inserts, updates, and deletes, which require X-locks, are not allowed. Once the S-lock
is removed, that is, once there is again only one transaction accessing the data,
inserts, updates, and deletes are permissible (X-locks are allowed). A transaction may
upgrade an S-lock to an X-lock only if there are no other transactions holding S-locks
on the same data.
With an X-lock, only one transaction at a time can use or lock the data. X-Locks are
applied to data on inserts, updates, and deletes. While the X-lock is in effect, other
transactions cannot access the data for any purpose. Transactions cannot place an Xlock on data while an S-lock on the data is in effect for some other transaction.
Transactions place U-locks when they have retrieved data with a clear intent to
update. This is usually done internally by SQLBase in order to avoid deadlock
situations, such as when building a result set for a UPDATE ... WHERE command
that will affect multiple rows. Only one U-lock can be held on a page at one time,
similar to X-locks, but this single U-lock can coexist with S-locks, unlike X-locks.
However, when the transaction holding the U-lock wishes to perform the update,
which requires upgrade of the U-lock to an X-lock, then the rules for X-locks apply.
This means that any S-locks that may have been coexisting (placed either before or
after the U-lock was set) with the U-lock must be released prior to the upgrade to an
X-lock taking place. These characteristics of U-locks make them less susceptible to
deadlock situations, as will be seen later in the discussions on isolation levels and
index locking.
The U-lock is the one lock in SQLBase that can be set explicitly. Whenever a
transaction issues a SELECT ... FOR UPDATE OF command, U-locks are placed on
the pages of the result set. If the program later issues an UPDATE ... WHERE
CURRENT OF command referring to the cursor of the SELECT, these U-locks are
upgraded to X-locks for the remainder of the transaction. Any U-locks that are still in
place are allowed to downgrade to S-locks when the cursor under which the SELECT
command was executed is used to prepare and execute a new statement (depending on
the isolation level in effect).
Advanced Topics Guide
Lock types and scope
Lock compatibilities
The following table summarizes the capabilities of each particular type of lock to
coexist with the other types.
Lock attempted to place
Existing lock on page
Note: *While S-locks and U-locks may co-exist when they are placed, the S-locks will have to
be released before the U-lock may be upgraded to an X-lock.
Locking level
SQLBase places locks at the page level. If the record being accessed fills most of the
page, only that record is locked. Otherwise, a group that is composed of the record
and its adjacent records, which together make up a page, is locked. A page can
contain any sort of data, including indexes and system catalog data (detailed
descriptions of the various SQLBase page types can be found in Chapter 8, Database
Locking scope
The scope of the locks implicitly placed by SQLBase cover all the data structures that
the optimizer determines are required to execute each command within a transaction.
When processing a SELECT statement, the optimizer will lock those index leaf pages
that are required to build the result set, if an index is chosen as a component of the
access path. Also, all data pages read during the processing of the SELECT statement
will have S-locks placed upon them. Often, contention between transactions occurs
not on the data pages but within the index structures due to the potential proliferation
of X-locks within the index. When S-locks are released is a function of the
transaction’s isolation level, but is also affected when transactions use cursor context
preservation. On the other hand, X-locks are always held until either a commit or
rollback is issued, regardless of the transaction’s isolation level or the use of cursor
context preservation.
Advanced Topics Guide
Concurrency Control
Cursor context preservation
Normally, all of a transaction’s locks are released by a COMMIT command. If the
transaction has turned cursor context preservation on (commonly abbreviated CCP)
for one or more cursors, through setting the SQLPPCX flag to one, then the rules for
releasing locks change somewhat. Specifically, those cursors’ result sets and
associated S-locks are maintained across the COMMIT command, allowing the
transaction to perform update operations and commit those results without losing its
current place in the result sets of open cursors. Also, for transactions running in the
RL isolation level, and not performing DDL commands, result sets will be kept intact
across the execution of a ROLLBACK command. This means that following the
COMMIT command for a transaction with CCP on for some cursors, locks will still
be held (as required by the transaction’s isolation mode) for those result sets
associated with SQL SELECT commands of those cursors.
The CCP flag allows programs to browse large result sets, perform updates based on
the result sets, commit or rollback those updates as desired, and then continue their
browsing without having to reestablish the result set. This makes these transactions
very effective in a multi-user environment, since the commit allows locks
(particularly X-locks) to be freed-up and which other transactions may be waiting on,
while the CCP feature allows the transaction to avoid the overhead of building the
same result set multiple times.
Index locking
In the case of an INSERT or DELETE, X-locks will be placed not only on the page
containing the row which is the subject of the command, but also on at least one page
for every index that exists on the table. This same case holds true for UPDATE
commands specifying columns which are the subject of one or more indexes.
The index page that will definitely be X-locked is the page containing the leaf entry
pointing to the subject record. This is the only page in the index which will have a
permanent X-lock placed on it, this X-lock will not be released until the transaction
commits or ends. If index structure adjustments need to be made, such as splitting
index records, then additional X-locks will be taken as well. These additional locks
will be taken on those nodes within the index structure that have to be updated to
perform a split or the movement of a successor key up the tree. These X-locks will be
released once SQLBase completes the operation on the index tree; they will not be
held through the life of the transaction (in contrast to the X-lock on the affected leaf
SQLBase performs the following steps when performing a DML INSERT operation
which must modify an index (both DELETE and UPDATE function similarly and
therefore are not explained):
Advanced Topics Guide
Lock types and scope
As SQLBase traverses the index tree, it places a temporary U-lock on each
node it accesses. This starts with the root node and progresses downward.
When a node within the tree is traversed, it is checked to see if it is full. If it
is not full, then all the temporary U-locks on nodes higher up in the tree are
released, not including the current node. This is done because any index split
that may be required could not effect any higher nodes than the current, nonfull node, since this node can accommodate another entry without splitting.
When SQLBase reaches the leaf node and it is not full, then no split will be
required and all temporary U-locks on higher level nodes can be immediately
released. If the leaf is full, then a split is required and all higher level tree
nodes that need to be adjusted as a result of the operation will have their Ulocks upgraded to temporary X-locks. Of course, the leaf node itself will
always be the subject of a permanent X-lock.
SQLBase performs the index updates, both to the leaf nodes as well as any
nodes further up the tree that may be affected.
The permanent X-lock on the leaf page is always held until the transaction
terminates with either a COMMIT or a ROLLBACK. All temporary X-locks
on higher level pages within the index structure can now be released.
The most important aspect is that some time during the performance of a DML
modification statement that affects an index, there will be a temporary U-lock on the
root page of the index. Hopefully this lock will be in place for only a very short time,
depending on the condition of the index tree and the insertion pattern followed by the
application. However, if some other transaction should come along and want to
perform an equivalent operation, they will have to wait in order to place their own
temporary U-lock on the root page as well. Therefore, any application environment
that is very update-transaction intensive should attempt to eliminate all but the most
critical indexes on those tables that are used by most or all transactions. Each index
placed on a table that is inserted, deleted, or modified at a rapid rate, will have the
effect of placing a single-threaded contention point into each transaction that
performs these operations.
Lock implementation
The mechanism used by SQLBase to implement locks is through an in-storage hash
table. This storage construct is used because it is efficient in storing and retrieving
tables with a large number of entries. When a new lock is needed, a previously
allocated, but currently free lock entry may be re-used. If no lock entries are available
for reuse, then additional lock entries may be dynamically added to the table in
groups of 100 locks at a time. These allocations can be added to the table until the
maximum number of locks specified in the LOCKS keyword of the SQL.INI file is
reached, or until the server has no more memory to allocate. Each lock entry in the
Advanced Topics Guide
Concurrency Control
table represents either a single page or a range of pages, and contains the following
The group number owning the page, or page range, that is locked. As
explained in Chapter 8, this number uniquely identifies one object within the
database, such as a table or index.
The logical page number of the page that is locked, or the starting logical
page number of a page range that is locked. Logical page numbers are
assigned sequentially within a group starting from one.
The logical page number of the end of the range which is locked.
The type of lock that is held, either S, U, or X. This is prefixed with a ‘T’ if
the lock is temporary, such as TS, TU, or TX.
The transaction identifier of the owner of the lock.
Whenever SQLBase is performing a read or a write operation, the lock table is
checked for the group and page on which the operation is performed. If no conflicting
lock type exists in the table, then a new lock is added on behalf of the current
transaction before the operation proceeds. In this way, SQLBase ensures that locks
are set prior to performing the operations they control, and that no lock conflicts exist
that could undermine the integrity of the system.
Isolation levels
The protocols that are used by SQLBase to perform locking operations on behalf of a
transaction are called isolation levels. These define how long locks will be held and
how SQLBase will pass data to the transaction via the message buffer. The four
available isolation levels are: Read Only (RO), Repeatable Read (RR), Cursor
Stability (CS), and Release Locks (RL). These are often referred to by their
abbreviations. The isolation level is selected by a transaction when it first opens a
cursor to SQLBase, and remains in effect for the life of the transaction. The isolation
level that is selected by the transaction applies to all cursors that will be used by the
transaction. Changing isolation levels in the middle of a transaction has the effect of
causing an implicit COMMIT and starting a new transaction with the new isolation
level. All of the transaction’s cursors will now operate under this new isolation level.
Also, having cursor context preservation on will not preserve the result sets of any
cursors, nor their locks, when changing a transaction’s isolation level in midstream.
13-10 Advanced Topics Guide
Isolation levels
Using isolation levels effectively
Repeatable Read and Cursor Stability deadlock avoidance
While the Repeatable Read (RR) and Cursor Stability (CS) isolation levels offer the
greatest levels of data consistency available to transactions that perform updates, their
low level of concurrency (especially RR) leads many application developers to avoid
them. These levels may be appropriate choices, however, for those systems that
demand a high level of data consistency, due to the nature of their application. This is
particularly true of the CS level when used in an intensive transaction processing
environment where few rows are created or updated by each transaction execution.
The use of CS for these cases allows a guarantee of consistency without severely
impacting performance provided that some guidelines are used in order to minimize
the occurrence of deadlock situations. One drawback in using the CS mode that
should be kept in mind is the increase in network traffic that may be caused by the
single-row buffering used by CS to respond to FETCH commands. The following
guidelines, when used with the CS isolation mode, will minimize deadlock situations:
All transactions should acquire their lockable resources (table rows) in the
same order. An example would be to read the CUSTOMER table, then the
ORDER table, then the ITEM table in that sequence. By conforming to an
ordered specification for reading and updating tables, different transactions
can avoid deadlocks that are caused by contending for locks between unlike
Transactions which perform updates of rows returned from a SELECT result
set should use the FOR UPDATE OF subclause of the SELECT in order to
place a U-lock at the time of the SELECT. When they perform the actual
UPDATE command, they then use the WHERE CURRENT OF subclause
specifying the cursor that holds the result set of the SELECT. When
transactions follow this protocol, they can avoid deadlocks between like
transactions competing for locks on the same tables. A secondary transaction
that accesses the same page as another will have to wait until the first finishes
before it can place its own U-lock, since only one U-lock may exist on a page
at one time.
The use of these two guidelines can greatly reduce the possibility of transactions
terminating due to true deadlock situations. Depending on the processing time of the
transactions, however, there may be time-outs due to waits for locks to be released by
their owners.
Release Locks isolation level
Locking strategies which are “pessimistic” prevent the situation from occurring where
lost updates or uncommitted dependency problems could arise. For example, as
explained above, an S-lock is placed on a data base row (in actuality the data base
Advanced Topics Guide
Concurrency Control
page) as soon as a transaction reads the data. At this point, SQLBase will not allow
any other transaction to update the row so that lost updates are guaranteed not to
occur. However, even though multiple transactions read the same data base row/page,
only one transaction may actually choose to update the row. This transaction is
prevented from doing so because one of the other transactions might update the row.
SQLBase offers an advanced “optimistic” isolation level which most multi-user
applications will choose to decrease the number of SQL timeout and deadlock errors.
SQLBase offers the isolation level Release Locks, which allows conditions to develop
where lost updates and uncommitted dependency problems might occur but provides
the programmer with features to detect and avoid these problems when they actually
occur. Each data base row in SQLBase is assigned a unique ROWID when the row is
initially inserted and this field is changed every time the row is updated and
committed. The programmer can track the value of ROWID and determine when a
row has actually been deleted or updated by another user.
Warning: Using Release Locks (RL) isolation level with programs which perform database
updating requires each program using the RL feature to conform to the same method of
detecting and handling cross transaction updates, such as the technique shown here. Failure of
all programs to conform will likely result in lost updates and uncommitted dependency
problems that will proliferate throughout the database.
For example, with the Release Locks isolation level, the programmer would issue the
following command to select a row from a customer table:
INTO:SavedRowID, :IntoVar1, :IntoVar2, ... :IntoVarN
The following SQL update command would be used when the programmer needed to
update this row:
UPDATE CUSTOMER SET COL1 = '<value>' ...
If another user had actually updated or deleted this row since this program selected it,
the ROWID would change such that the SQL update command would fail and the
data base server would return the -806 SQL error code. The programmer can then test
for the -806 error and take appropriate action (such as advising the user that some
other user has either updated or deleted this customer since they read it, and ask if the
user wants the program to re-read the customer record in order to get a fresh copy).
Read-Only isolation level
The Read-Only (RO) isolation level in SQLBase is a special isolation level that is
available for use only to transactions that perform no updates to the database. This
mean no INSERT, UPDATE, or DELETE DML statements are allowed. The RO level
achieves data consistency without setting any S-locks through a technique called
13-12 Advanced Topics Guide
Isolation levels
multiversion concurrency control. This allows any transaction running in RO to read
all of its data as it existed at the time of the SELECT statement execution, without
having S-locks placed on the result set of the SELECT.
The way SQLBase handles this is by maintaining a history file of database page
images. This history file contains the “before-update” image of database pages for a
certain timeframe. Each of these images is a version of the database page at some
prior point in time. When an RO transaction retrieves a page from the database,
SQLBase compares the timestamp of the last page update with the timestamp of the
SELECT statement being processed. If the database page has the more recent
timestamp, then SQLBase goes to the history file to retrieve a version of the page as it
existed prior to the update, and passes the rows from this page back to the
transaction’s program.
The Read-Only isolation level offers the highest possible levels of concurrency and
data consistency for read only transactions, but there is a significant overhead cost
associated with the ongoing work required to maintain the history file and the
multiple versions of database pages contained therein. This overhead can be as high
as 30% for databases supporting OLTP type applications, with many users executing
rapid transactions. The availability of the RO isolation level is controlled by the DBA
through settings in each database’s control page, as well as globally through a server
flag. These settings are probably best turned off (RO disabled, which is the default
setting) unless a significant need for this high level of data consistency exists, in
which case they may be turned on only for those periods of time when RO
transactions are known to be running.
Selecting isolation levels
Selecting isolation levels is an activity that should be performed by a person
knowledgable of the overall application characteristics. The selection of an
appropriate isolation level requires knowledge of the overall application design and
usage patterns that may be beyond the scope of every project team member’s
knowledge. Furthermore, the use of isolation levels should be standardized across the
entire application since the choice of one transaction’s program can affect the other
transactions with which it will be interleaved.
Comparing isolation levels
When designing multi-user database applications, two conflicting factors must be
considered before choosing an isolation level:
Data Consistency. Data should be consistent when accessed by multiple
User Concurrency. Multiple users should be able to access the same data
Advanced Topics Guide
Concurrency Control
The highest degree of data consistency occurs at the expense of user concurrency. For
example, one locking protocol may ensure that lost updates or uncommitted
dependency problems will not occur. However, this may result in many timeout and
deadlock errors. A SQL timeout error occurs when a given SQL command waits a
program-defined number of seconds (the default is 360 seconds, which is 6 minutes)
for a lock to be released by another program so that it may access the data base page.
A deadlock occurs when user A is waiting for row 1 which user B has, and user B is
waiting for row 2 which user A has. In this situation, SQLBase will sense this
deadlock and abort one of the requests thereby allowing one of the users to continue.
Da ta Consistency
U ser Concurrency
Figure 13C - An ordered approximation of isolation levels along the data consistency
vs. user concurrency continuum. Note that Read Only (RO) occupies both ends of the
spectrum only because it disallows any update operations and thereby avoids X-locks
Criteria for choosing an isolation level
As we have seen, SQLBase provides several isolation levels. The choice of an
isolation level is determined by evaluating the following criteria:
Buffers: How critical is performance for this application? If performance is
critical, and meeting performance expectations is going to be difficult,
choose an isolation level that sends one buffer of rows at a time to the client
(meaning it fills the entire input message buffer before sending any data).
Avoid using the Cursor Stability (CS) isolation level, which sends one row at
a time.
Contention: Will there be a great deal of user contention, where many users
will be attempting to access the same data simultaneously? If so, do not use
an isolation level that places S-locks on all read data until a commit/rollback.
Update: Will the users be performing inserts, updates, and/or deletes? If so,
the isolation level must allow these actions.
Consistency: Is a consistent view of data across the application important? If
transactions must know about updates the moment they are made by other
transactions, or must guarantee that data they have read will maintain the
13-14 Advanced Topics Guide
Isolation levels
same values until the transaction ends, then select an isolation level that holds
read locks until commit or end.
The following table may be used to help determine which isolation level is
appropriate for a specific database application based on responses to the previous
criteria description.
Criteria 4
Data Consistency
Locks maintained for
duration of transaction,
more than one page
may be locked.
High data consistency,
low user concurrency.
S-locks held only for
rows on current page.
Medium data
consistency, medium
user concurrency.
No locks, no inserts,
updates, or deletes
High data consistency,
highest user
No held S-locks, Slocks released at endof-fetch, X-locks held
until commit/rollback.
Low data consistency,
high user concurrency.
Common isolation level scenarios
Finally, there are some common rules-of-thumb that can be followed in the great
majority of cases when selecting an isolation level for a transaction. While these
represent common practice by SQLBase users, they do not substitute for knowledge
of the lock process and how the various isolation levels work. These are simply a set
of heuristic rules that can speed up the decision process when the transaction has no
unusual characteristics that may warrant deviation from normal practice.
For transactions which are executed in a multi-user manner throughout the
day and may perform a moderate amount of update activity, the RL or CS
isolation levels are the best choice. These levels provide the greatest
concurrency when used in conjunction with good programming practices,
such as committing frequently. The choice between these two isolation levels
is often made on the following basis:
When the application will always be used solely with SQLBase, as
opposed to another RDBMS, then the RL isolation level should be used.
Advanced Topics Guide
Concurrency Control
If there is a possibility that the application may be ported to another
RDBMS, then the CS isolation level should be used.
These choices are based on the fact that few RDBMSs provide such an
optimistic locking protocol as RL (which makes the assumption that
many rows read by a transaction will not need to be updated), but the CS
isolation level is a standard feature in most RDBMSs and will usually be
available. The CS protocol is the most commonly used for transaction
processing systems utilizing IBM’s DB2 mainframe RDBMS system.
Examples of transactions falling into this category are order entry, confirm
shipment, and receive inventory transactions in an order processing system.
For update transactions that require their entire view of the data to be
consistent throughout their execution, the RR isolation level should be used.
Because of the high number of locks that may be held, these transactions
should be scheduled to execute at times when the system is relatively idle,
otherwise a large number of time-outs and deadlocks may result. These
transactions may also include rather elaborate commit and restart logic in
order to free up as many locks as possible that are not needed.
Examples of these types of transactions, (again in an order processing
application), would be processing physical inventory count, purging aged
orders, or applying updated product prices.
Transactions which meet the preceding definition of requiring a consistent
view of their data throughout their definition, may also be candidates for the
use of the RO isolation level, provided they do not perform any database
modifications whatsoever. Since the RO mode sets no locks, the concern of
causing time-outs and deadlocks is eliminated, and the transaction could be
scheduled at any time regardless of system workload considerations. The
downside, however, is that SQLBase must perform the work of maintaining
history files for the storage of the prior versions of database rows, causing
both processing overhead and more disk utilization. Of course, this drawback
is not as significant if update transactions run rarely, or not at all, since the
need to build rows in the history file would occur infrequently.
An example of a transaction in this category would be a year-to-date sales
report for an order processing system.
Transactions which do not easily fit into any of the preceding categories should be
evaluated carefully as unique cases. Consideration should be given to the
transaction’s usage frequency, including transaction volume and times, the program’s
commit and restart logic, and the characteristics of the workload that will be
executing concurrently. This analysis should yield an appropriate isolation level for
any transaction. If problems are encountered when the transaction begins normal
operations, the isolation level may need to be reconsidered.
13-16 Advanced Topics Guide
Isolation levels
Remember, when going from a restrictive isolation mode to a more relaxed mode
(such as from RR to RL), there may be programming modifications required to make
the transaction behave properly in the new isolation level. An example is checking
ROWIDs for concurrent updates by transactions that are executing in RL mode.
Application designers should be cautious when migrating to more relaxed isolation
levels to ensure that data consistency is not sacrificed to the point where the
application can no longer be considered reliable by its users.
Advanced Topics Guide
Advanced Topics Guide
Chapter 14
Logging and Recovery
In this chapter we:
Review the various types of failures that the SQLBase server is susceptible to.
Examine the way SQLBase uses the log files to minimize the possibility of
database damage occurring during a failure.
Evaluate the backup and recovery techniques that you can implement in order
for the SQLBase safeguards to function effectively.
Advanced Topics Guide
Logging and Recovery
There are many kinds of failures to which computer systems are susceptible. In order
to protect the database against the effects of these failures, SQLBase contains a
Recovery Manager (RM) software component which is responsible for ensuring that
the database is always in a consistent and valid state. This consistent state is defined
simply as having all of the effects of committed transactions applied and none of the
effects of uncommitted transactions applied. The RM achieves its goals primarily
through the use of a log file. This file contains the physical images, or history, of
database changes made by transactions over time. For this reason, the log file used by
SQLBase is a physical log file, hereafter called the log file. Since the RM component
of the database server provides the recovery method for all database transactions,
regardless of their program characteristics, SQLBase is said to contain centralized
This chapter will first discuss the broad types of failures which SQLBase protects
against, then the mechanisms used to provide this protection.
Failure types
There are many possible failures that a computer system is susceptible to. Some of
these failures are caused by faulty programs that place erroneous data into the
database (but which commit nonetheless), or the entry of flawed data into the system
by application users who are unaware of the damage they are doing. Unfortunately,
these are examples of errors which cannot be recovered automatically, since any
DBMS must trust that the execution of program requests which are semantically
correct will place correct results in the database. As will be seen, though, some of
these types of errors can be recovered from by using the same techniques as are
applicable for media failures. The three basic types of system failures which the RM
provides centralized recovery from are transaction failures, system failures, and
media failures.
Transaction failures
Transaction failures are aborts of transactions that result in a performing rollback for
the transaction. The transaction may abort at its own request (by issuing the
ROLLBACK command), or due to some other cause, such as a client workstation
failure or a network failure. Transaction aborts may also be issued by the SQLBase
server in some situations, such as to break a deadlock situation between two
transactions, one of which must be aborted in order for the other to proceed. Also, a
transaction failure may affect one transaction only, or some number of transactions at
Advanced Topics Guide
Failure types
SQLBase provides transaction integrity, which prevents these types of failures from
corrupting the database. The mechanism used to accomplish this is by undo
processing from the log file. Following the undo, all database updates performed by
the transaction are no longer present.
System failures
System failures cause loss of the server’s volatile storage (RAM memory). The term
“system” as applied to this type of failure refers to the program processing
environment that is created by the operating system (OS/2 or other SQLBasesupported environment) and consists of the resources and control blocks that allow for
task ownership and dispatching functions. These failures may be due to causes such
as an abnormal end (abend) in the operating system environment (such as NetWare),
an operator inadvertently cancelling the server program, or power failures, among
others. The net effect of these types of failures is that either the SQLBase server task
suddenly disappears, or it discovers that it can no longer continue processing and
must request an abend. When the SQLBase server is shutdown normally (when no
users are connected to the database), or a database is de-installed from the server
through the DEINSTALL command, the Database Control Block (DCB) page for the
database is flagged to indicate a normal shutdown took place. This is called a
“graceful shutdown”.
SQLBase provides for system integrity through the implementation of a restart feature
which prevents these failures from affecting the database. Restart is invoked every
time a database is first connected to after the server is started, where it examines the
DCB page of the database to determine what led to the prior shutdown. If it finds that
the flag in the DCB indicates no recovery is required, then it knows that the system
had been terminated gracefully on its last execution. In this case, restart can be
bypassed and the connection established immediately. If the flag in the DCB indicates
that recovery is required, then restart must perform its processing prior to allowing the
connection to the database to be established. Because of the dependency of the restart
logic on the contents of checkpoint records, this process is explained in detail in the
checkpoint section later in this chapter.
Media failures
Media failures cause the loss of some portion of the server’s non-volatile storage (disk
drive). These failures may be caused by environmental effects on the disk storage
environment or through defects in the storage device. Also, magnetic media are
inherently subject to a limited life because eventually the bonding material holding
the oxide on the platter (in the case of a disk) or the film (in the case of tape) will fail
and the oxide will begin to wear away. When this happens, minor errors will begin to
occur. If these errors are observed, the data may still be sufficiently readable to be
copied successfully to another device. Otherwise, the oxide will eventually degrade to
the point that the device will become unusable. This generally occurs to one isolated
Advanced Topics Guide
Logging and Recovery
storage device at a time. Multiple disk drives rarely fail together, unless some drastic
environmental problem occurs, such as a power surge or shock damage. These
potential disasters are the justification for developing and maintaining a disaster
recovery plan, which involves backing critical data up onto tape or disks that may be
stored off-site from the computer hardware (which is highly recommended for critical
computer installations, but is beyond the scope of this chapter).
Sites can establish procedures using SQLBase’s backup and recovery features which
allow the impact of media failures on the database to be minimized. Through periodic
backup of the database and log files the capability is provided to restore to any
particular point in time that is desired. In order to recover to a specific point in time,
the required backup is simply restored. Any log images that need to be processed to
that restored database in order to reach the desired point in time are then processed
through a SQLTalk command called ROLLFORWARD. The most important
characteristic of this process is that the application of redo log images in the
rollforward is an “all or nothing” choice. There is no option to bypass either some
transactions on a log file, or some log files, or skip over a particular time gap. For
whatever point in time is chosen to recover to, all log images prior to that point in
time must be applied. SQLBase will not allow you to attempt to do otherwise, since
this would probably cause severe database corruption which would most likely be
Recovery mechanisms
The SQLBase recovery manager relies primarily on two mechanisms for
accomplishing its centralized recovery objectives. The first of these is the log file,
which provides the information required for undo, redo, and checkpoint processing.
The other is the set of utilities which provide for backup and recovery of databases.
The log file
As mentioned in the introduction, the log file contains the physical images, or history,
of database changes made by transactions over time, and is therefore called a physical
log file. Another important characteristic of the log file is that it exists on stable
storage, as opposed to being stored in volatile storage. Stable storage is an area that
retains memory across system failures, typically a disk drive. Volatile storage is the
RAM memory used by the running system, which is subject to loss of content due to
a variety of types of failures. Since the log resides on stable storage, it becomes a
dependable repository of historical information about the contents of the database and
the way these have been altered over time by the execution of transactions. This
Advanced Topics Guide
Recovery mechanisms
information is the key to the Recovery Manager’s ability to perform automatic
recovery from a variety of failure types.
Figure 14A - There are two basic types of logs, active and archive. Archive logs may
be moved to tape or other media through use of the BACKUP LOGS SQLTalk
command when LOGBACKUP is on.
The log files used by SQLBase can be divided into two subtypes, the active and
archive logs. Together, these two log types make up the complete database log file.
The active log is the log that is currently being written to by the server; this is a disk
file whose status is “open”. For this reason, the active log is pinned to the disk and can
only be processed by the server. Only one active log can exist at any single point in
time. Whenever the server is shutdown gracefully, or a database is de-installed, all
“after images” on the log that have not yet been externalized to the database are
written (they are contained in dirty cache buffers), and the active log becomes an
archive log. When the server restarts, or the database is re-installed, a new active log
file is allocated.
There are usually many archive logs, however. These are simply old active logs that
have filled up and are no longer needed by the server. When LOGBACKUP is
enabled, these logs should be backed up to some stable storage media that is separate
from where active log and database files are stored in order to safeguard their
integrity. Of course, prior to this backup, some of the archive logs will still exist as
files in the same directory or DBAREA as the active log, while waiting for the next
log backup.
Note that log file names are formed by using a sequential number for the first part of
the name, allowing for log files from 1 to 99,999,999 to exist. Since this number is
reset to 1 whenever the database is reorganized, there is little likelihood of it ever
rolling over. Appended to this number is an extension type of LOG, which identifies
the file as a log file.
Log files are stored in a subdirectory of the LOGDIR directory that is named the same
as the database which owns the log file. This convention allows for identification of
Advanced Topics Guide
Logging and Recovery
the owning database only so long as the log file is in a subdirectory that applies only
to that database. Merging the log files of multiple databases into one subdirectory will
eliminate the ability to ascertain which log files belong to which databases. If these
log files are subsequently needed to ROLLFORWARD, neither the DBA nor
SQLBase will be able to ascertain which log files to make available to the rollforward
process. For this reason, DBAs should take care not to allow the log files of multiple
database to become merged in a single directory.
Pinned log files
Some archive logs may also be pinned to the disk if they may still be needed by the
server for transaction recovery or restart purposes. The reasons a log may be pinned
to disk are:
The log is active, which means the server is currently writing to it, as
explained previously.
If there are records on it from uncommitted transactions, then it may be
needed for later recovery. Should one of these transactions issue a
ROLLBACK, then the before row images stored on this log would be
required for the undo operation.
If it contains a system checkpoint that is either the most recent or the next to
most recent checkpoint, then it may be needed for later restart.
The log has not yet been backed up, and LOGBACKUP is in effect.
Figure 14B - The active log, 37.log in this case, is always pinned. The prior log,
36.log, is pinned because it contains log images for active task 49. The 35.log is
pinned because it contains a checkpoint record that could be needed for restart. Note
that 36.log would be pinned even without transaction 49, due to 35.log being pinned.
Advanced Topics Guide
Recovery mechanisms
SQLBase periodically reevaluates the status of all pinned log files and deletes those
that no longer meet one of these criteria. Log file status evaluations are performed
whenever one of the following events occurs:
At log rollover, usually when the current active log fills up and a new log
must be created. This action can also be forced to occur even when the current
log is not full through execution of the SQLTalk RELEASE LOG command.
Note that some SQLBase commands issue a RELEASE LOG implicitly, such
as the BACKUP SNAPSHOT command.
When log files are backed up through the SQLTalk BACKUP LOG
command, or the equivalent API call.
When the system’s “next log to backup” parameter is reset manually through
the use of the SQLTalk SET NEXTLOG command, or the equivalent API
When logging for the database stops because either the server was shutdown
or the SQLTalk SET RECOVERY OFF command is executed.
These are the only events that cause SQLBase to consider deleting pinned log files.
Performing an evaluation of the logs’ pinned status whenever a transaction performs a
commit or rollback would cause excessive overhead and slow down transactions, and
is therefore not done by SQLBase.
Record types
The SQLBase logging manager is an integral part of the SQLBase DBMS. Whenever
a database program executes, it creates before and after images for those portions of
database rows that are modified in some manner. These modifications are performed
through execution of the INSERT, DELETE, or UPDATE DML commands. Other
commands which update the system tables in the database, including DDL and DCL
commands, also cause the before and after row images of the catalog tables they
update to be logged. Also written to the log at the appropriate times are control
records indicating when the program connected to the database (start record), and
when it issued a COMMIT or ROLLBACK command (commit and abort records).
These log records allow SQLBase to maintain transaction integrity and perform point
in time recovery operations. In addition, SQLBase writes a number of system control
Advanced Topics Guide
Logging and Recovery
records to the log file that allow it to maintain the integrity of the page allocation
mechanism and enhance the ability to perform system restarts.
Figure 14C - The composition of the log file. Each log file contains many variable
length records such as LR1 through LR5, above. Each record may be one of several
types. An expansion of the image fragments for the data page log record type is
As shown in figure 14C, each log file physically contains many variable length
records. Each record consists of a header, a variable portion, and a trailer. The log
record header is 34 bytes in length, and contains the following fields:
Type: the format type of the log record, such as those explained below.
Length: the total length of the log record.
Timestamp: the point in time that the log record was written.
Transaction: the ID of the transaction which caused the log record to be
Chaining information: a pointer connection two related log records.
Each of log record is of one particular type, such as:
Start transaction
This log record is written whenever a program first connects to the database,
thereby starting a transaction. There is no variable portion to the start
transaction log record type.
Advanced Topics Guide
Recovery mechanisms
Commit transaction
This log record is written whenever a program issues a COMMIT command.
There is no variable portion to the commit transaction log record type.
Abort transaction
This log record is written whenever a program issues a ROLLBACK
command, following completion of the undo operation. There is no variable
portion to the abort transaction log record type.
Data page modification
Any change made to a table’s data rows will cause this record to be written.
Note that only those columns that actually change in the row will be logged.
These changes are made through the execution of the DML commands
INSERT, UPDATE, and DELETE (an exception to this case is system
catalog tables, which have changes made to them through the execution of
DDL commands). The variable portion of this record type has two major
Before fragment
The before fragment is written whenever a previously existing row is
modified through execution of the UPDATE or DELETE DML
commands. The fragment has a 5 byte header, followed by the before
images of the changed row columns, and ends with a 2 byte trailer.
After fragment
The after fragment is written whenever a DML command such as
INSERT or UPDATE causes changed data to remain on a database page.
This fragment has a 5 byte header, followed by the after images of the
changed row columns, and ends with a 2 byte trailer.
Index page modification
Whenever a program issues a DML command that INSERTs or DELETEs a
row a table that has an index associated with it, or UPDATEs columns which
are the subjects of an index, then some change will be required to be made to
the index structure. These changes are logged as index page modifications,
with a format similar to the database page change log record type. Each time
a node of the index tree is modified, an index page modification log record is
written which contains the before and after fragments of the changed portion
of the index node. Note that CLUSTERED HASHED indexes are not subject
to this logging activity, since they have no index tree structure.
Advanced Topics Guide
Logging and Recovery
A checkpoint log record is written during each system checkpoint. The
variable portion of the record contains a 22 byte entry for each active
transaction connected to the database at the time the checkpoint was
Page allocation
This log record type is written whenever pages are allocated from the
database. There are usually three of these log records written for each page
allocation action, the largest of which contains the after image of the
allocated page’s header and footer, which is about 60 bytes long. The other
two log records record the changes made to other pages which are chained to
the newly allocated page, each of these log records contains about 8 bytes of
variable information.
Long VARCHAR object
These log records are written whenever a column whose type is LONG
VARCHAR is inserted, modified, or deleted. The log record contains one
page that has been allocated to store the LONG VARCHAR column in the
variable portion of the log record. If the column is longer than one page, then
multiple log records will be written, each containing one of the pages
allocated for that column.
Undo and redo logic
The system functions that read the log and apply images from it are undo and redo.
The undo process (also called rollback) is used when before images are to be restored
to the database, while the redo (also called rollforward) performs the restoration of
after images to the database.
The mechanism used to accomplish the restoration of before images to the database
for failed transactions is by undo processing from the log file. This process starts at
the end of the active log and reads backward from that point. As the log is read, each
before image of an update is restored to the database. This continues until the start
control record of the transaction is reached, at which time the abort termination
record can be written at the end of the log file, and the transaction terminated. The
transaction’s locks are held until transaction termination to prevent other transactions
from accessing the rows that are being restored to the database; these locks are finally
released when the abort record is written to the log. The strict serialization that is
forced upon transactions through the use of database locks assures the integrity of the
undo process.
The mechanism used to accomplish the restoration of after images to the database
when the updates of committed transactions must be reapplied is through redo
processing. This process involves reading forward through the log. As the log is read,
14-10 Advanced Topics Guide
Recovery mechanisms
each after image of a database change is checked against the database to ensure that it
was successfully written, and if not, then the redo process writes it to the database. If
the end of the log is encountered, any active transactions that remain (those for which
no termination record has been found) are the subject of undo processing, as if they
had been the subjects of a transaction failure.
One important characteristic of the undo and redo processes in SQLBase is that they
must be capable of recovering the database to its committed state regardless of the
status of updated pages in cache. Cache is the volatile system memory used as a buffer
area for database pages. A large quantity of cache memory can greatly accelerate the
performance of SQLBase by decreasing the number of read and write operations that
must wait for external disk accesses to complete. The maintenance of the cache
buffers is the responsibility of the Cache Manager (CM), which performs fetch
operations to return the contents of cache to SQLBase, and flush operations to dispose
of the contents of cache pages. In SQLBase, system efficiency is enhanced by
allowing the CM to make the majority of decisions regarding the timing of cache
operations, such as when to flush pages and when to write “dirty” pages to disk
(“dirty” cache pages are those pages which have been updated in the buffer but not yet
written to disk). In order for the undo and redo processes to work successfully in this
environment, SQLBase must follow two rules regarding the interplay between the
CM and external storage:
Undo rule (also called the write ahead log protocol). This rule states that
before an uncommitted value may be written to the external database
(DASD), its prior value must be saved in stable storage (the log). This rule
ensures that no previously committed row can be overwritten by an
uncommitted row without its contents being saved in stable storage for
possible undo processing. This means that the log record containing a before
image of an updated row must be written prior to the CM’s flushing of the
dirty cache buffer containing the updated page. The CM enforces this rule by
placing the log record pointer within the dirty cache buffer, and not allowing
flushes of cache buffers whose associated log records have not yet been
Redo rule. When processing a transaction’s commit, all write operations must
be completed to stable storage. This can consist of either the log file or the
external database. This rule ensures that committed values are present in
some form of stable storage, and not just within the contents of dirty cache
buffers. SQLBase flushes log buffers at commit time to ensure compliance
with this rule.
Advanced Topics Guide
Logging and Recovery
Checkpointing and restart
One problem faced during system restart is how to determine what log records need
to be processed out of the voluminous contents of the total log file. The entire log file
cannot be scanned and processed due to the enormous amount of time this would
take. Simply examining the current log file also is inadequate, since this may take
quite a while if the file is large, and some uncommitted transactions may have been
active on prior log files. Because fast restarts are an important component of overall
system availability time, SQLBase employs a strategy called fuzzy checkpointing in
order to speed this process up.
In general, the activity of checkpointing is the writing of certain control information
to the log on a periodic basis during normal processing in order to speed up the restart
process. This information is summarized from the DBMS’s internal control blocks
and consists of lists of transactions and the states that they are currently in, so that the
restart process can rapidly ascertain what processes need to be performed, and for
which transactions. These lists, organized around the possible states of a transaction
(such as active, committed, or aborted), are organized chronologically and also
contain markers indicating when previous system checkpoints were taken.
When a checkpoint is actually being created, the DBMS must reach some consistent
state in order for the externalized control blocks to be stable as well. During the time
required to reach this consistent state, the DBMS must block the work (or suspend
processing) of all active transactions, otherwise the point of time of the consistent
state would always be “after this next operation” and would never actually be
reached. The amount of time that transactions are blocked while this process happens
depends on the amount of consistency that the checkpoint activity needs to achieve
prior to writing a checkpoint record. Whatever inconsistencies the checkpoint process
allows to remain must be resolved by the restart module whenever it is required to
run. This is a classic trade-off situation, with higher levels of consistency taking
longer to achieve at checkpoint time but yielding greater benefits in terms of
simplified restart processing. Of course, since the DBMS spends much more time
performing routine processing than restarting (hopefully!), getting carried away with
too much consistency during checkpointing has the penalty of causing transaction
delays during the creation of each checkpoint record for no subsequent benefit
The fuzzy checkpointing scheme used by SQLBase requires a minimum amount of
system consistency to be achieved, and therefore only blocks active transactions for a
very brief interval when writing the checkpoint record. The benefits during restart are
still substantial, though, since most of the time only the log file records between two
checkpoints and the end of the log must be traversed in order to complete the restart.
These two checkpoints are the last and the second to the last (or penultimate)
checkpoint in the log. The log files containing these two checkpoints will always be
14-12 Advanced Topics Guide
Recovery mechanisms
pinned to disk to ensure their availability. The following tasks are performed by
SQLBase when creating a checkpoint:
Write to the log a list of transactions which are currently active, or processing.
Flush all dirty cache buffers which have not been written since before the
prior checkpoint. If the server has been busy, many of these cache buffers will
have been written during the time interval since the last checkpoint because
the server needs to re-use cache pages based on the least recently used (LRU)
algorithm. If this is the case, then few cache pages should require flushing at
this checkpoint, allowing the checkpoint to be processed faster. Alternatively,
if the server has been largely inactive since the last checkpoint, there may be
numerous pages containing data updated prior to that checkpoint that must
now be written to disk. If this is the case, however, the server will likely still
be relatively inactive and therefore the work associated with writing these
buffers should not have an significant impact on current transactions. The use
of this fuzzy checkpointing feature allows SQLBase to perform checkpoint
processing with a low likelihood of negatively affecting performance.
When SQLBase starts, and discovers that a restart is required, the following steps are
The next to the latest checkpoint record in the active log is located. This is called
the penultimate checkpoint record.
Redo processes all transactions following the penultimate checkpoint marker.
During the redo processing, the last checkpoint record in the log will be
encountered. This checkpoint contains a list of all active transactions at that point
in time. Redo uses this to track all transactions from this point on in the log.
When a transaction ends during redo processing, it is removed from the list of
active transactions.
Eventually, redo will encounter the end of the log file.
Undo now reads the log backwards from the end to process all entries in the active
transaction list. Since these transactions had not been committed prior to the end
of the log, they must be rolled back.
When undo encounters the start control record of the last transaction entry in the
active list, recover is complete and the database is available for new connections.
This process ensures that the database is returned to a consistent state, both in terms
of containing data only from committed transactions and not losing the contents of
any write-delayed cache buffers.
Advanced Topics Guide
Logging and Recovery
Backup and recovery
SQLBase’s backup and recovery features allow sites to establish procedures which
minimize the impact of media failures on the database. This is made possible by
“point in time” recovery, which allows the database to be restored to any prior point
in time, thereby eliminating any updates or failure events that occurred between that
selected point in time and the present. This also allows recovery from database
corruption caused by program logic errors, bad data, or faulty recovery processes. By
periodically backing up the database and log files and saving them, the backup and
recovery capability is provided to restore to any particular point in time that is
desired, so long as two conditions are met:
There must be an existing database backup that was created either at, or prior
to, the point in time where the database will be restored to.
There must be a continual series of log files (either backed up or still on disk)
from the database backup point in time up until the point in time being
recovered to.
The tools necessary to perform these tasks are provided with SQLBase. Other backup
programs and devices may be substituted if higher capacity or quicker backup is
desired through the use of products such as high capacity tape devices.
Since the recovery process requires identification and coordination of the required
backup and log files, sites should implement some form of labeling and tracking
procedures for whatever media they decide to store these files on. The recovery
process depends on reliable and rapid identification of these files in order to be
completed in a minimum amount of time. Also, the frequency of these backups and
log off-loads plays a key part in the site’s recovery procedures.
Considerations for partitioned databases
There are special considerations that must be allowed for when performing backup
and recovery operations on partitioned databases. This is because the extent
allocation information for a partitioned database resides in the SQLBase server’s
MAIN database. Without the MAIN database, the server will be unable to access the
information in the partitioned database. Therefore, the MAIN database must always
be kept in synchronization with the partitioned database it controls. Whenever
activity against a partitioned database is controlled by SQLBase, this level of
synchronization is guaranteed. Only when using utilities external to the SQLBase
software must the DBA be concerned about ensuring that the state of the MAIN
database and the site’s partitioned database reflect the same point in time.
Since partitioning is usually performed only for large databases, off-line backup
techniques are often appropriate. The SQLBase on-line backup utilities may be used
with a partitioned database, but either adequate disk space will have to be available
14-14 Advanced Topics Guide
Recovery mechanisms
for creation of the backup image file or the operating system has to provide a
technique to redirect the output directly to a tape device.
When using off-line backup techniques to backup partitioned databases, extra care
must be taken to insure that the MAIN database remains synchronized with the
partitioned database. This means that during backup, the server must be shut down,
and the partitioned database and the MAIN database must both be backed up prior to
restarting the server. When performing a restore made through this scenario, the
server must be shut down, and both the partitioned and MAIN databases restored to
the matching set of backups. Failure to maintain this parallelism between off-line
backup and restore techniques will cause SQLBase to be unable to determine the
proper location of the database suballocations within the DBAREAs assigned to the
partitioned database, which will result in unrecoverable database corruption.
When database backup and recovery procedures are implemented using the SQLTalk
BACKUP, RESTORE, and ROLLFORWARD commands, SQLBase modifies the
contents of the MAIN database as necessary to ensure that they accurately reflect the
current state of the suballocations within the partitioned database’s DBAREAs.
Therefore, the DBA can be unconcerned about explicitly managing the backup and
recovery of the MAIN database in this situation. Of course, backing up the MAIN
database might still be considered as a part of a disaster recovery plan that allows a
site to restore server operations on another machine that does not have an initial
installation of SQLBase software.
Advanced Topics Guide
Advanced Topics Guide
Chapter 15
Introduction to Query
In this chapter we:
Follow the evolution of data access languages from their procedural roots to
their modern declarative implementation represented by the SQL language
used by SQLBase.
Examine the various categories of query optimizers that exist in relational
database management systems.
Advanced Topics Guide
Introduction to Query Optimization
Data access languages
Before we can address query optimization, it is necessary to review data access
languages. You use a data access language to store data into and retrieve data from a
database in a database system. Two fundamental different types of data access
languages exist today: procedural (or navigational) and declarative.
Procedural data access languages
Before the introduction of SQL database systems, most database systems (for
example, IDMS, IMS and dBase) were based on procedural, or navigational, data
access languages. Procedural data access languages require the programmer to code
the programming logic necessary to navigate the physical data structures to identify
and access the required data. For example, with IDMS, programmers must write
navigational logic to FIND CALC a specific record and OBTAIN NEXT within
specific sets. Similarly, in dBase, programmers USE a specific index.
By referring to the physical data structures in the application programs, the
application programs become dependent upon them. Application programs then
require updating if changes occur to the physical structures. For example, if an index
is deleted in dBase, all application programs that use that index must be modified.
This dependency between the application programs and the physical data structures
significantly increases the cost of application development and maintenance. During
the development of large complex computer systems, it is quite common to discover
inadequacies in the physical database design during the programming phase of the
project. To remedy these inadequacies, the database administrator must implement
changes to the physical structures.
These schema changes in turn impact all existing programs which reference the
altered physical structures. These programs must be updated to reflect the new
schema. If the program relied on some altered physical structure to navigate the
database and other programming logic was based on this navigation, a significant
amount of re-coding may be required. Likewise, maintenance of existing systems
often results in schema changes and therefore this dependency also drives up the cost
of maintaining an application.
In addition, procedural data access languages are often proprietary. Consequently,
application programs become closely bound to the particular database systems on
which they were developed. This binding of application programs to particular
database systems severely limits the ability to migrate these application programs to
improved and lower cost database systems.
Advanced Topics Guide
Data access languages
Declarative data access languages
With the introduction of SQL and relational database systems, it became possible to
sever the link between application programs and the physical data structures.
Declarative access languages only specify what data is needed by the application
program, leaving it to the database system to determine how to navigate the physical
structures to access the required data. SQL is an example of a declarative data access
Decoupling the application programs and the physical data structures results in
several significant benefits:
You can respond to changing data requirements without impacting
existing application programs. For example, if existing indexes become
obsolete, you are free to drop existing indexes and create new ones without
impacting existing programs. The new indexes created may result in either an
increase or decrease in program performance; however, you are assured that
the existing programs will execute without error (assuming they PREPARE
their SQL prior to execution, some RDBMSs such as SQLBase may also
automatically recompile stored SQL access plans as well). This is not true
with procedural data access languages that abort program execution
whenever physical structures are either not found or have been altered.
Application program complexity is reduced. The database system, not the
programmer, determines how to navigate the physical structures. This frees
the programmer for other tasks, since this aspect of programming is often the
most complex aspect of the program’s logic.
Application program error is reduced. The complexities of data access
often result in program error unless the programmer is highly skilled and
extremely careful. The primary advantage of a computer is the ability to
execute simple instructions at a high rate of speed and without error.
Consequently, database systems, in general, out-perform programmers when
determining how to navigate physical structures to access required data.
The component of SQL database systems that determines how to navigate the
physical structures to access the required data is called the query optimizer. The
navigational logic to access the required data is the access path and the process used
by the query optimizer to determine the access path is called query optimization.
Advanced Topics Guide
Introduction to Query Optimization
Query optimization
The query optimization process determines the access path for all types of SQL DML
statements. However, the SQL SELECT statement (a query) presents the biggest
challenge to access path selection; therefore, the process is commonly referred to as
query optimization rather than data access optimization. Further, it is worth noting
that the term query optimization is a misnomer in that there is no guarantee that the
query optimization process will produce an optimal access path. A more appropriate
term might be query improvement. For example, the best possible access path given
the known costs. However, we use the standard term to conform to convention.
Early relational database systems processed queries in a simple and direct manner
without attempting to optimize either the query itself or the access path. This resulted
in unacceptably long query processing times, and led to the belief among some
application programmers that relational database systems were not practical for real
world applications. To speed query processing times, a great deal of research and
testing was carried out to discover ways to increase the efficiency of query
processing. Query optimization can thus be defined as the sum of all techniques that
are employed to improve the efficiency of processing queries.
Syntax optimization
The first progress in optimizing queries was in finding ways to restate the query so
that it produced the same result, but was more efficient to process. For example,
consider the following query:
The most direct way to process this query is:
Form the Cartesian product of the vendor and product tables.
Restrict the resulting table to rows that meet the WHERE condition.
Project the three columns named in the SELECT clause.
If the vendor table has 50 rows and the product table has 1000 rows, then forming the
Cartesian product would require 50,050 rows to be read and written. Restricting the
result would require 50,000 more rows to be read from the resulting product, and if
20 rows meet the WHERE condition, 20 rows would be written. Projecting the result
would then require 20 additional reads and writes. Processing the query in this
manner would require 100,090 rows to be read and written.
Advanced Topics Guide
Query optimization
However, one of the algebraic rules that apply to SQL is that
(table_a JOIN table_b) WHERE restriction on table_a
is equivalent to:
(table_a WHERE restriction on table_a) JOIN table_b
This means that restrictions on a table can be performed as early as possible to reduce
the number of rows that must be processed further. Applying this rule to the example
above, we can process the query as follows:
Restrict the rows in the part table to those for which VENDOR_CODE = “100”.
This would require 1000 reads and 20 writes.
Perform the join of the result to the vendor table in step 1. This would require 20
reads of the intermediate result above, which presumably could be kept in
memory. Joining the intermediate result with the vendor table would then require
100 additional reads, and 20 writes.
Processing the query would require only 1120 reads and 40 writes to produce an
equivalent result. The transformation described in this example is called syntax
optimization. There are many formulas for making transformations, and all modern
query optimizers use transformation rules to restate a query into a more efficient
form. The important point to realize is that these automatic transformations free you
from having to try many equivalent forms of the same query in search of the one form
that yields the best performance, because the optimizer is usually transforming all of
these queries into the same final form.
Rule-based optimization
As progress was being made in improving the processing of queries, other efforts to
improve table access were taking place. These improvements included indexing
techniques and hashing. However, these improvements added complexity to query
processing. For example, if a table had indexes on three different columns, then any
one of these indexes could be used to access the table (along with sequential table
In addition, many new algorithms to join tables began to appear. The two most basic
join algorithms are:
Nested Loop Join - A row is read from the first table, called the outer table,
and then each row in the second table, called the inner table, is read as a
candidate for the join. Then the second row in the first table is read, and again
each of the rows in the inner table is read again, and so on until all rows in the
first table are read. If the first table has M rows, and the second table has N
rows, then M x N rows are read.
Advanced Topics Guide
Introduction to Query Optimization
Merge Join - This join method assumes that the two tables are sorted (or
indexed) so that rows are read in order of the column (or columns) on which
they will be joined. This allows the join to be performed by reading rows
from each table and comparing the join column values until a match is found.
In this way, the join is completed with one pass through each table.
Further, join operations are both commutative and associative; consequently, it is
theoretically possible to perform joins in any order. For example, all of the following
statements are equivalent:
(table_a JOIN table_b) JOIN table_c
table_a JOIN (table_b JOIN table_c)
(table_a JOIN table_c) JOIN table_b
However, different access paths, join algorithms, and join orders may result in
significantly different performance. Consequently, when joining several tables each
with multiple indexes, there exists literally hundreds of possible combinations of join
orders, join algorithms and access paths to select from. Each produce the same result
but with varying performance characteristics.
The first method of dealing with this complexity was to establish heuristic rules for
selecting between access paths and join methods, which is called rule-based
optimization. With this approach, weights and preferences are assigned to the
alternatives based on principles that are generally true. Using these weights and
preferences, the query optimizer works through the possible execution plans until it
arrives at the best plan based on its rules. Some of the rules used by this type of
optimizer are based on the placement of variable tokens such as table or column
names within the query’s syntax structure, when these names are moved a dramatic
difference in the performance of the query can sometimes be realized. For this reason,
rule-base optimizers are said to be syntax sensitive, and one of the methods for tuning
this type of database system involves moving tokens to different positions inside a
Rule-based optimization provides satisfactory system performance in those situations
when the heuristics are accurate. However, often the general rules are inaccurate. To
detect these situations, the query optimizer must consider the characteristics of the
data such as:
The number of rows in the table
The range and distribution of values for a given column
The row width and consequently the number of rows per physical page on
The height of an index
The number of leaf pages in an index
Advanced Topics Guide
Query optimization
These data characteristics can greatly affect the efficiency of processing a query. This
resulted in the next type of optimization.
Cost-based optimization
Cost-based optimization is similar to rule-based optimization, except that cost-based
optimizers use statistical information to select the most efficient execution plan. The
cost of each alternative execution plan is estimated using statistics such as the number
of rows in a table and the number and distribution of the values in a table column. The
cost formulas typically consider the amount of I/O and CPU instructions required to
perform the execution plan. The statistics are stored in the system catalog and
maintained by the database system.
To understand how statistics can make a difference, consider the following query:
If an index exists on the STATE column, a rule-based optimizer would use this index
to process the query. However, if ninety percent of the rows in the customer table have
FL in the STATE column, then using the index is actually slower than simply
processing the table sequentially. A cost-based optimizer, on the other hand, would
detect that using the index would not be advantageous.
The cost-based optimization approach today represents the state-of-the-art in query
optimization techniques. Most high quality relational database systems, including
SQLBase, employ a cost-based optimizer.
Query optimization steps
While the query optimizers of modern relational database systems differ in
sophistication and complexity, they all follow the same basic steps in performing
query optimization:
Parsing - The optimizer first breaks the query up into its component parts, checks
for syntax errors, and then converts the query to an internal representation for
further processing.
Conversion - The optimizer next applies rules for converting the query into a
format that is syntactically optimal.
Develop Alternatives - Once the query has undergone syntax optimization, the
optimizer next develops alternatives for processing the query.
Create Execution Plan - Finally, the optimizer selects the best execution plan for
the query, either by following a set of rules (rule-based optimization) or by
computing the costs for each alternative (cost-based optimization).
Advanced Topics Guide
Introduction to Query Optimization
Because steps 1 and 2 take place without regard to the actual data in the tables used, it
is not necessary to repeat these steps if the query needs to be recompiled. Therefore,
most optimizers will store the result of step 2 and use it again when it reoptimizes the
query at a later time.
Advanced Topics Guide
Advanced Topics Guide
Chapter 16
Overview of the
SQLBase Query
In this chapter we:
Review basic relational operators implemented within the SQL data access
Examine the physical operations used by SQLBase to perform the relational
Learn how to view SQL statements as query trees, which simplifies the
processing steps required for their execution.
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
In the next three chapters we examine the query optimizer contained within
SQLBase. The SQLBase query optimizer (often simply called the optimizer) is a
specific implementation of a cost-based optimizer, meaning its decisions on access
path selections are based on the estimates of potential costs associated with executing
a particular access plan.
The basis of these cost estimates are the statistics contained within the SQLBase
catalog and control areas within the database. The advantages of this approach are
that the access plan decisions made by SQLBase can be made with very recent
information concerning the actual database contents. Just how recent and accurate
this information is depends on the procedures you follow in administering your
This chapter reviews the basic relational operators implemented by SQL followed by
an overview of the techniques that SQLBase may use when actually executing these
operations. This set of techniques, called physical operators, is the set of actions the
optimizer considers when building access plans.
While we summarize the actions taken to perform each physical operator, we will
also briefly summarize the major factors comprising the cost of the operator (which is
usually I/O). Finally, the structure of an access plan is explained, with an emphasis on
the heuristics used by the optimizer when building a set of access plans on which to
perform cost analysis.
Throughout the next three chapters, examples that illustrate optimizer operations
generally use the SQL SELECT statement. You should keep in mind, however, that
all SQL Data Manipulation Statements (DML) are processed by the optimizer. This
includes not just the SELECT statement, but the INSERT, UPDATE, and DELETE
statements as well.
The steps involved in optimizing these other statements are virtually identical to those
of the SELECT, the only difference being operations used to perform locking and
logging during update operations. These exceptions are covered in Concurrency
Control and Logging and Recovery chapters. The benefits of using the SELECT
statement for examples is that the SELECT offers many more relational operations,
such as a join, which exercise a much wider variety of optimizer processes than other
DML statements.
A secondary advantage of using SELECT statements for examples is that the
execution of a SELECT statement creates a result set which visually shows the final
output of SQLBase’s processing. This contrasts to the other DML statements, which
change the state of the database, but these changes are not visible externally except
through the return of error codes or other environmental variables. Only a subsequent
Advanced Topics Guide
Relational operations
SELECT statement will reveal the changes which were made to the contents of the
database by the execution of one of these DML commands.
Relational operations
The basic operators of the relational algebra were defined by E. F. Codd when he
published his landmark paper delineating the relational model access language in
1972. The operations defined in this relational algebra form the logical operators that
must be processed by the optimizer. These logical operators are implicit in the syntax
of SQL DML statements submitted to SQLBase for processing.
The following relational algebra operations have one or more relations as inputs, and
one relation as output. A relation is simply a table of some degree n, where n is the
number of attributes (or columns) in the table. These operations are broken down into
two classes: traditional set operations and special relational operators. While the
traditional set operations are sufficient to process most multiple relation queries, the
special relational operators specify the desired results more efficiently. This last class
of operators is the basis of the SQL language, and is examined in more detail
throughout the remainder of this chapter. The traditional set operations are defined
here to provide background information.
Traditional set operations
An important distinction of the traditional set operators is that (with the
exception of the Cartesian product) the relations used as input to these
operations must be of the same number of columns, with each matching
attribute defined on the same domain. This means that each table must have
the same number of columns, and that each column pair of any position (for
example, column 1 of each table) must be defined with the exact same data
type and scale.
The union of two relations is all rows belong to either or both relations.
It can be formed by appending one table to the other and eliminating
duplicate rows.
The intersection of two relations consists of the rows which appear in
both relations. Any row which appears in only one table, but not the
other, is not a member of the intersection set.
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
The difference between two relations is all rows which belong to the first
table but not to the second. Note that this operation is not commutative
like the other traditional set operators, meaning that A – B ≠ B – A .
Cartesian product
The Cartesian product of two relations is the table resulting from
concatenating each row of one table with every row of the other table.
The table resulting from this operation contains as many rows as the
product of the rows in the input tables. This means that the Cartesian
product of a 15 row table and a 50 row table contains 750 rows. As
mentioned previously, this is the only traditional set operation where the
format of the two tables can differ.
Special relational operations
SQL generally uses these operations more than the traditional set operations:
The projection operation limits the columns of a table that are referenced
in a SQL statement.
A select limits the result set to only those rows that meet the
qualifications. The selection criteria compares one or more columns of
the table to either one or a set of constants or expressions.
The join operation creates a result set by concatenating rows together
from multiple tables. These concatenated rows must also meet the
specified join criteria, which is a comparison between one or more
columns of the tables involved. These comparisons differ from the select
operation in that they compare columns of different tables.
Because of the central role played by these special relational operators in SQL
statements, we will now examine them in more detail.
Advanced Topics Guide
Relational operations
When a SELECT references specific columns of one or more tables, only those
columns are returned:
The alternative to performing a projection is to use the asterisk (*) wild card notation
which causes all columns to be included:
Note that many sites have standards prohibiting this notation from being used in code,
since changing the number of columns in the table can have an adverse affect on the
program containing the statement.
The select is the row equivalent of the column projection operation. In SQL, the
WHERE clause specifies a selection by naming columns in comparison expressions
with either constants, that consist of one or a set of values, or expressions that evaluate
to a single value or set of values. The following are examples of the select operation:
(TOTAL_AMT > 1000);
Each of the comparisons contained within the WHERE clause is called a predicate.
Some SQL statements, such as the last example shown here, contain more than one
predicate. When the operation specified in a predicate is performed on a row, it is
called applying the predicate. If the predicate evaluates to TRUE, it is said to be
satisfied, otherwise it is not satisfied. When a statement has more than one predicate,
they must be connected by one of the Boolean operators AND or OR. When all of the
predicates of a statement are connected by AND, they are said to be AND-connected,
or conjunctive. When they are all connected by OR, they are OR-connected, or
disjunctive. This terminology of predicates plays an important role in how they are
used by the query optimizer, as we shall see later.
Note that the form of the WHERE clause that compares columns of two different
tables is not a select operation, but rather a specification for a join operation, which is
covered next.
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
A join creates its result set from two or more tables in a similar manner as the
Cartesian product mentioned earlier; it concatenates each row of one table with every
row of the other table. The difference between the Cartesian product and the join is
that only rows meeting some specified join criteria are eligible for inclusion in the
result set. This is logically equivalent to building the Cartesian product, then
performing a select operation against it. However, the join allows for a more efficient
operation with most query optimizers because they can eliminate rows from the result
set without first forming the entire Cartesian product as an intermediate result.
To perform a join, use the SQL FROM and WHERE clauses. The FROM clause must
name each input table to the join. The WHERE clause specifies the join criteria, by
comparing one or more columns of one table with columns of another table. The
predicates used as join criteria determine how many rows of the Cartesian product
will be bypassed in processing the join. The more restrictive the join criteria are, the
more efficient the join will be when compared to the Cartesian product (note that the
Cartesian product is formed by not specifying any join criteria at all). Here are some
examples of joins:
In the second example, the ‘C’, ‘O’, and ‘P’ letters allow you to qualify column
names with their correct tables without having to spell out the entire table name. In
SQL, these are called correlation variables. Since most column names need to be
qualified they are often used in joins due to the presence of multiple table names in
the FROM clause.
Two important characteristics of the join operation are that it is both commutative and
associative in nature:
A JOIN B = B JOIN A (commutative)
A JOIN (B JOIN C) = (A JOIN B) JOIN C (associative)
(A JOIN B) JOIN C = B JOIN (A JOIN C) (both)
The effect of these properties is that when processing a join of more than two tables,
the optimizer may select any sequence of two table joins to complete the operation.
This allows the optimizer to treat a join of any number of tables as a series of two
table joins, which avoids having to have special physical operations available to join
any specific number of tables.
The type of comparison performed is often used to further refine the name of the join.
These sub-classifications of the join are briefly covered below.
Advanced Topics Guide
Relational operations
The most common example of the join is the equijoin, which uses the ‘=’ operator.
This version of the join operation is used whenever business relationships are used to
combine information from two tables. These joins occur by setting a primary key in
the parent table equal to the foreign key in the child table.
Natural join
This is a particular form of equijoin where all columns of the two tables are included,
with the exception of one set of the key columns, which are duplicated across the
tables. This is a full equijoin between two tables without any projection except for the
elimination of the duplication of the key columns.
The semijoin operation is equivalent to an equijoin with a subsequent projection that
only includes the columns of one of the join tables. This is useful when rows of a table
are needed that meet a selection criteria that is defined in terms of membership within
another table’s foreign key column. An example would be the set of all products
which have been ordered during January of 1993:
(O.ORD_DATE BETWEEN JAN-1-1993 AND JAN-31-1993);
The :equijoin is also useful in a distributed environment when the two tables to be
joined are at different nodes on the network.
The outerjoin allows the non-participating rows of one of the joined tables to be
included in the result set. The reason these rows are termed non-participating is
because they contain key values which are not referenced by any rows of the other
table. For example, a row in the product set which contains a product number that has
never been ordered by any customer would be a non-participating row in a join of the
product table with the order table. In SQLBase, these rows could be included in the
result set anyway through the specification of the outerjoin operator, ‘+’, on the order
table key as in the following example:
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
A selfjoin is an equijoin of a table with itself. This is also called a ‘recursive join’.
Note that in this case, you must assign correlation variables in order to avoid
ambiguous column references, as in the following example which lists the names of
all employees and their assigned managers:
While aggregation was not originally specified as a relational operation, its inclusion
as a standardized capability in SQL served to make it a commonplace operation. The
purpose of aggregation is to represent a table of information through some derived
statistic, such as the sum or mean (average) of a set of numbers.
SQL supports two types of aggregation, the simplest is scalar aggregation, which
derives a single output from a column of a relation. Scalar aggregation in SQL is
performed through the inclusion of an aggregate function in a query that does not
have any GROUP BY clause. Examples of scalar aggregation include:
The first query produces the total salary paid to all employees, while the second query
tells how many employees are named ‘DAVE’. The distinguishing characteristic of
these queries is the presence of some aggregate function in the SELECT list which is
allowed to range over the entire result set of the table access (after performing the
project and select operations).
The second type of aggregation supported by SQL is function aggregation. This type
of aggregation uses the same aggregate functions as scalar aggregation. The
difference between the two is that function aggregation always creates another table
as output. This contrasts with scalar aggregation, which simply returns a single value
as output.
The SQL syntax of a statement performing function aggregation always includes the
GROUP BY clause. The columns named in the GROUP BY clause become, in effect,
the key to the new table that is output from the statement. Examples of function
aggregation are:
Advanced Topics Guide
SQLBase physical operators
Here the first query lists the department and average salary for each department, while
the second lists the birth month and number of employee birthdays in each month
throughout the year. In each of these statements, the aggregation function only ranges
over the rows of the input table that have equal values in the columns specified in the
GROUP BY clause.
SQLBase physical operators
When SQLBase executes a query plan, it reads the plan and performs each specified
step in turn. Each of these steps tells SQLBase what operation to perform at that step,
and what inputs and outputs are required. When executing query plans, SQLBase uses
physical operators. These are different than logical operators such as SQL statements
which specify relational operations that must be performed. For every possibly logical
operator, there is at least one, and possible many, physical operators that allow
SQLBase to execute the operation in an efficient manner.
Every physical operator has either one or two inputs, depending on the nature of the
operation. It also has one output table (which of course may consist of only a single
row, or even no rows). This output table may be a result set which is the final output of
the query, or it may be an intermediate table which will be used as an input table to an
operation later in the query plan. One query plan can contain many of these
intermediate tables. They can be stored within SQLBase’s cache memory, or on disk
in a temporary file. When a step in the query plan is finished executing, the temporary
input tables it uses are either freed from memory or deleted from disk.
The sort operation is performed whenever a table must be placed into a specific
sequence according to one or more columns, which are collectively called the sort key.
This physical operator has one table as input, and one table as output. The basic
technique SQLBase uses for sorting is the postman sort, which is a radix sort.
The aggregation operator also has one table as input, and one table as output. The
SQLBase physical aggregation operator performs a serial scan of the qualifying data
rows, with the aggregate function calculation being performed for each row. When
SQLBase performs simple scalar aggregation, the input rows can be in any sequence.
However, when aggregate functions are being evaluated through SQL statements
containing GROUP BY clauses, the input table to the aggregation operation must be
in sequence by the columns specified in the GROUP BY. This prerequisite is satisfied
by the query optimizer through the placement of either a sort operator prior to the
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
aggregation, or through the use of a table access technique which returns table rows
in a specific sequence (discussed later in this chapter).
Disk access operation
This group of physical operators is responsible for returning rows from a single table.
These rows can then either be processed by other steps in the query plan, or constitute
the final result set, depending on the SQL statement being processed. Disk access
operators are entirely responsible for the access of data from disk, however.
Therefore, one of these operators always appears as the first step of a query plan, in
order to perform the access of data to whatever table the optimizer decided should be
processed first.
Table scan
This is the simplest technique for accessing a base table. Each data page associated
with the table is read in succession. For each page, the rows on that page are extracted
for processing. Note that this extraction may require accessing additional pages in the
form of either extent pages or overflow pages (in the case of hashed tables). In the
absence of any extent or overflow pages, the number of disk accesses required for a
table scan is equal to the number of data pages assigned to the table, including pages
required for any LONG VARCHAR columns that may be specified.
Index leaf scan
This access technique is used whenever a specific sequence is desired for the rows of
a table, and an index matches this sequence, but there are no predicates specified for
the table that are present in the index. In other words, this access technique can
substitute for a sort if the index sequence matches the desired table sequence.
The basis for this technique is the B+Tree modification to the basic B-Tree, which
links together the leaf pages of an index. This list of linked leaf pages is called the
sequence set of the index. It is traversed in key sequence, with each index node’s rows
accessed as the node is processed. This means that the worst case I/O count is the
number of leaf pages in the index plus the number of rows in the table. This is
sometimes improved on, however, by the table being either fully or partially clustered
in the index sequence. We will see how the optimizer accounts for this possibility in
the next chapter.
Matching index scan
The matching index scan uses the full capabilities of SQLBase’s B+Tree index
structures. This technique is useful whenever a SQL statement only requires a portion
of a table to be processed, based on predicates that match the columns present in the
index. Since the generic capabilities of B+Tree indexes are exploited by SQLBase,
this access technique can use any index which has predicate columns in the leftmost
key positions. The optimizer generally uses this technique if the predicates are
16-10 Advanced Topics Guide
SQLBase physical operators
restrictive enough to make the extra I/Os for the index worthwhile. Also, if an
inequality predicate is used for some index column, this technique uses the index tree
to find the starting leaf, and then scans the sequence set forward or backward from
that point.
The number of I/Os incurred by this access technique is one for each level in the
index, plus one for each index node that is accessed, plus one for every row that must
be retrieved. The optimizer estimates costs associated with this access technique
using an index statistic called the selectivity factor, which is explained in detail in the
next chapter. Briefly, the selectivity factor describes the expected number of rows in
the table that would satisfy a predicate.
Hash access
A table which has a clustered hashed index placed on it may be accessed via this
index whenever all of the columns contained within the index are used in equality
predicates in an SQL statement. As explained in Chapter 10, Hashed Indexes, the
hashing technique lacks any generic or scanning capabilities. This is why the hashed
index key must be fully specified with the equality predicate in the SQL statement.
Whenever the restrictive conditions placed on the use of the hash access technique are
met, this is probably the quickest possible alternative for table access and will likely
be selected by the optimizer. The I/O cost of this technique is one I/O for each row, or
set of rows, which has the key specified in the predicates. Extra I/O costs may be
incurred for searching a chain of overflow pages if the hashed table has experienced a
large number of collisions, and this condition may cause the optimizer to reject the
hash access technique.
Join operations
The relational join is one of the most frequently executed operations of any relational
database. It can also be extremely time consuming due to its data intensive nature.
This is because the most simplistic implementation of a join requires every row of the
one table to be compared to every row of another table. In order to execute joins as
quickly and efficiently as possible, SQLBase selects from among a number of
alternative techniques. All of these operators accept two tables as input, and produce a
single table as output. These input and output tables can be any combination of
database tables, temporary tables, or final result sets.
Terminology that is common to all join algorithms is the naming of the two tables
involved. Since one row of a table must be compared to all rows of another table, the
table for which a row is held is called the outer table. The table which is scanned
while a row is held from the outer table is called the inner table. These terms first
originated with the loop join algorithm, where they are especially appropriate, as we
shall see next.
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
Simple loop join
This is the most straightforward method for joining two tables, and can be used for
any join criteria. It can also be the worst performing join operation, especially if both
input tables are large. One advantage of the simple loop join is that it can be used in
any case, regardless of the type of join criteria specified or the absence of index
structures. Also, there are cases where the simple loop join can be the best performing
technique, particularly if one of the input tables can fit entirely into cache memory.
The loop join algorithm obtains a row from the outer table, then scans the entire inner
table to find the rows which meet the join criteria. Whenever the criteria are met, the
rows are appended and output. When the end of the inner table is reached, a new row
is obtained from the outer table, and the scan of the inner table is started again.
Due to the nested loop structure of this technique, the worst case for the number of I/
O operations is the number of data pages in the inner table multiplied by the number
of data pages in the outer table. This can be improved on dramatically if the inner
table is small enough to fit into cache memory, when the number of I/Os drops to the
sum of the data pages in the two tables. For this reason, SQLBase always chooses the
smaller of the two tables to be the inner table when cache is sufficient to hold the
table, at which point this join technique may become less costly than other
The following two alternative access methods are enhancements to the basic loop
join. They could be used whenever a suitable index is present, in which case they can
speed up the join process greatly over the simple loop join.
Loop join with index
This variation on the nested loop join can be used when there is an index on one of
the tables that has the join columns in the high order positions of the key. When this is
the case, the optimizer uses the table with the index as inner table of the loop
algorithm, and uses the index to limit the search for rows matching the current row
being held from the outer table. This greatly decreases the number of I/Os needed to
process the inner table. The best case is when the inner table is clustered, and an
equijoin is being performed, in which case the number I/Os is equal to the sum of the
number of data pages in the two tables, plus the height of the index tree, plus the
number of rows in the outer table. The last two cost components account for the
overhead of accessing the index structure when joining row occurrences.
Loop join with hash index
This enhancement of the loop join technique can be applied when one of the tables to
be joined has a hash index placed on it. The other two prerequisites for this technique
are that the hash index must be placed on all of the columns of the join criteria (and
no others), and the join must be an equijoin. When these conditions are met, this join
technique can be very efficient. The worst case is when the number of I/Os is equal to
16-12 Advanced Topics Guide
SQLBase physical operators
the number of data pages in the outer table plus the number of rows in the inner table.
This only occurs if every row in the inner table is referenced by the outer table. When
not all rows of the inner table are referenced in the outer table, there is a reduction in
the number of I/Os.
Index merge join
This technique can be used when there are indexes present on the join criteria for both
tables. The basic algorithm is to scan the sequence set of the two indexes, and merge
rows from the table based on their join criteria being satisfied. For instance, when
processing an equijoin, first an entry is read from the sequence set of the outer table.
Then, an entry is read from the inner table’s sequence set. If the inner table’s join
column is less than the outer table, the scan of the inner table’s sequence set
continues. If the inner table’s join column is greater than the outer table’s, the outer
table’s sequence set is scanned until either equal or greater values are found. When
the join columns in the two indexes satisfy the join criteria, the rows of the tables are
read from the data pages, appended to each other, and written to the output table.
The I/O cost of this technique depends on how many rows of the tables meet the join
criteria. The best case is on an equijoin of primary keys, when the cost are equal to the
number of leaf nodes in the two sequence sets, plus the number of rows in the two
tables. If the tables are also clustered by the sequence sets, the number of I/Os to fetch
rows is reduced to the number of data pages allocated to the two tables plus the sum
of leaf pages on the two indexes.
Hash join
This technique can be used for equijoins only, and does not require any indexes on the
join criteria for the two tables. The algorithm first performs the setup phase, when it
scans the outer table and places each row into a hash table at a location determined by
applying a hash function to the join criteria. The smaller table is always chosen as the
outer table, in order to minimize the memory demands of the hash table.
In the second phase, called the probe, the inner table is scanned and the same hash
function used in the setup is applied to the join columns in order to locate a possible
match. Any rows found in the hash bucket from the first table are then examined. The
rows that satisfy the join criteria are appended to the inner table row and written to the
output table.
This can be a very rapid join technique, especially when the smaller table can be
hashed into a table fitting in memory. When this is the case, the I/O required is the
sum of the data pages of the two tables. Otherwise, additional I/O is required to
partition the hash table into segments, and manage those segments in a temporary file.
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
Imbedded operators
The two most common relational operators, select and project, are implemented
within the other physical operators. This allows the optimizer to perform these
operations at the earliest possible time in an access plan, which immediately reduces
the amount of data that the remainder of the plan must handle. The implementation of
the project operator is further enhanced by performing index-only access when
appropriate, as explained in the following paragraphs.
The select operation restricts access to rows that do not satisfy the predicate logic of
the query. These predicates are evaluated as early as possible when executing the
query, which is usually the first time the row containing the column named in the
predicate is accessed. Each file access physical operator in SQLBase accepts
predicates to perform this row filtering when accessing a table. If the predicates in the
query are conjunctive, or and-connected, a form of logic called short-circuit testing
may be performed. This technique allows the row to be discarded as soon as the first
unsatisfied predicate is found.
When the SQLBase optimizer estimates the cost of a query, the effect of the
predicates used for the select operation is taken into account through a statistic called
the selectivity factor. This figure represents the proportion of rows in a table that
would be expected to satisfy a particular predicate. The optimizer makes its estimate
of selectivity factor based on statistics associated with the table’s indexes and certain
heuristics concerning the predicate’s comparison operator. We will examine this
process in more detail in the next chapter.
The project operation is implemented within other physical operators by passing the
list of columns to be retrieved to any file access operator. That operator then only
includes those columns in its output table. Performing the project at the earliest
possible time in the query plan reduces the width of intermediate tables and result
sets, thereby saving space both in memory and temporary disk files. Note that one
side effect of performing the project at page access time is a possible reduction in I/
Os due to avoiding access to extent or LONG VARCHAR pages which do not contain
needed columns.
Index-only table access
Another benefit of building the project operation into the other physical operators is
the ability to perform index-only access. This occurs whenever the columns that are
specified for projection in the query are all present in one of the table’s indexes. That
index must also perform whatever predicate logic is required for the select operation.
When this happens, SQLBase’s optimizer realizes that the queries data requirements
16-14 Advanced Topics Guide
Query plan structure
can be satisfied without access to the table’s data pages. This can result in significant
I/O savings, depending on the number of rows which satisfy the predicates of the
Query plan structure
When the optimizer groups a set of physical operators together to execute a SQL
statement, the resulting specification is called a query plan.
An analogy to the programming task makes the role of the query plan easier to
understand. A programmer codes instructions in some high level computer language,
such as C, which is then compiled to produce an executable form of the programmer’s
code. This executable form uses machine instructions to talk to the CPU that runs the
program. When the CPU executes the program, it processes machine instructions
rather than the programmer’s original code.
Similarly, a programmer codes SQL statements, which are then processed by the
optimizer to produce the executable form of the programmer’s SQL, which is the
query plan. When SQLBase executes the SQL statement, it actually processes the
query plan, not the original SQL. Just as the C code has been translated into a more
rapidly executable form by the C compiler, the SQL statements are converted into a
form which the SQLBase engine can process quickly and efficiently.
Query trees
One way to view a query that provides for easier understanding is called the query
tree. This is a tree structure, where the final result of the query is at the top of the tree
(the root), and the database tables involved in the query are at the leaves. The
intermediate nodes of the tree, between the leaves and the root, represent physical
operations that are performed during the query execution. When the query plan is
executed, the sequence of node traversal is from bottom to top, and from left to right.
This is the sequence in which the operations are printed when the ‘SET PLANONLY’
option of SQLTalk is executed to print out a query plan for a given SQL statement.
The following example shows a query tree for a SQL statement that gets the names of
authors who published books in 1993:
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
proje ct
A .A N A M E
B .Y E A R ='1 9 93'
A .A N A M E =B. AN A ME
Figure 16A - A query tree for: Get names of authors who published
books in 1993.
Building query trees
When the optimizer builds query trees for cost estimation and possible final selection,
it uses some simple rules that determine the placement of various nodes in the tree.
These rules allow the optimizer to limit the number of query trees that are built for
further consideration by the cost estimation task. By limiting the possible access plan
alternatives in this way, the optimization process can execute more efficiently.
One of the important heuristics is to apply restrictions such as the select and project
operators as early as possible. This is because select and project operations can
reduce the size of the result table. In figure B, the optimizer improves the query tree
16-16 Advanced Topics Guide
Query plan structure
of Figure A by applying a select first to reduce the number of rows that are input to
the equijoin operation.
p ro je c t
B .Y EAR = '1 9 9 3 '
se le c t
Figure 16B - Using a heuristic rule to move the SELECT operator
down in the query tree in Figure A.
Furthermore, when multiple select operations are performed in a query against a
table, the ones with the lowest selectivity factor are performed lower in the tree, in
order to reduce the size of the intermediate results as rapidly as possible. Of course,
when predicates can be evaluated depends on when all the data referenced by the
predicate is available. SQLBase converts some predicates in order to evaluate them as
rapidly as possible, as we shall now see.
Predicate logic conversion
Whenever multiple predicates are specified using the relational select operation in a
query, an important heuristic is to convert them into an equivalent condition in
conjunctive form, which is a regrouping created by ANDing a set of predicates where
each predicate contains only OR operators. For instance, the clause:
WHERE COL1>5 OR (COL2<500 AND COL3>150)
can be rewritten as:
WHERE (COL1>5 OR COL2<500) AND (COL1>5 OR COL3>150)
Advanced Topics Guide
Overview of the SQLBase Query Optimizer
The conjunctive form is preferable since it can be evaluated with short-circuit
techniques. A set of predicates in conjunctive form evaluates to TRUE only if every
OR grouping evaluates to TRUE. It evaluates to FALSE if any grouping evaluates to
FALSE. In the short-circuit technique, if the result of evaluating any single OR
grouping is FALSE, there is no need to evaluate other conditions at all. Since the
number of comparison operations that are performed in a query have a direct impact
on the amount of CPU time needed, the short-circuit evaluation of predicates in
conjunctive form can save CPU cycles.
16-18 Advanced Topics Guide
Advanced Topics Guide
Chapter 17
SQLBase Optimizer
In this chapter we:
Discuss the details of how the SQLBase query optimizer performs access
plan analysis.
Summarize the statistical information available to the SQLBase query
optimizer to develop access plan cost estimates.
Explain how the query optimizer identifies possible access plans for
processing an SQL statement.
Review the two cost components SQLBase estimates for each access plan
using statistics about the database:
Total number of I/O operations.
Total CPU time required.
Advanced Topics Guide
SQLBase Optimizer Implementation
Database statistics
When the SQLBase query optimizer performs access plan selection, it uses statistics
about the database to estimate the costs associated with each candidate access plan.
These statistics, which are stored in the SQLBase system catalog tables, are
associated with the various tables and indexes contained within the database. Some of
these statistics are maintained dynamically, meaning that SQLBase keeps them
internally updated during normal server operations. The majority of statistics,
however, are maintained statically, meaning that they are only calculated and stored
when specific events occur.
The event that usually causes these statistics to be refreshed is the ‘UPDATE
STATISTICS’ SQL statement. This command examines a database’s tables and
indexes, calculates the required statistics, and then stores them in the appropriate
tables. Statistics are also created for an index or a table when it is initially created. If
no rows are present at this time, the statistics are set to default values.
Starting with SQLBase version 6.0, statistics can now also be updated directly by the
DBA, using the user modifiable statistics feature. In this feature, the access paths
chosen by the optimizer reflect the cost calculations that would be performed if the
database was populated in a manner that would yield the statistics placed in the
catalog by the DBA. In this way, the DBA can influence the optimizer into making
decisions that are based not on the current database contents, but rather on the
database as the DBA projects it will be populated at some point in the future. In order
to accurately calculate and store these statistics, the DBA must have a thorough
understanding of what these statistics mean.
Statistics that are dynamically maintained by SQLBase are used by the optimizer
whenever the DBA has not overridden the catalog statistics using the user modifiable
statistics feature. When you override the catalog statistics with your own calculated
values, the optimizer will them ignore the dynamic internal statistics in favor of your
catalog statistics.
Table statistics
The statistics associated with database tables are maintained in two separate system
catalog tables, each of which is described in the following paragraphs. All columns
not specifically mentioned as being maintained dynamically are static, and are
populated by running UPDATE STATISTICS or at CREATE time.
Advanced Topics Guide
Database statistics
Statistics in SYSTABLES
The following statistical information is maintained in these columns of the
SYSADM.SYSTABLES table in the system catalog:
This is the number of rows in the table. This statistic is maintained
dynamically in the table’s control page, but is only externalized to the
SYSTABLES column when the UPDATE STATISTICS command is
This is the total number of data pages in the table. This is also a dynamic
statistic that is externalized when UPDATE STATISTICS is run. Also, when
this column is maintained by the system, it will always be the sum of the next
two columns, ROWPAGECOUNT and LONGPAGECOUNT. If, on the
other hand, the DBA has set these statistics explicitly, then this relationship
may not hold true.
This is the number of base row pages occupied by the table, plus any extent
pages that may also be allocated. This is another dynamic statistic which is
only externalized on command.
This is the number of pages allocated to the table for the storage column
defined with the LONG VARCHAR data type. This is a dynamic statistic
which is only externalized on command.
This is the average contribution of data stored in base and extent row pages.
It excludes contribution from long pages.
This is the actual average length of rows in the table. This can differ
significantly from the defined row length since SQLBase stores all columns
as variable data regardless of the data type used in the column definition. Note
that this row length is only the length of the row as stored on base table and
extent pages, and therefore excludes all LONG VARCHAR columns.
This is the actual average length of all LONG VARCHAR columns stored in
the table. This column will contain zero if there are no columns of this data
type stored.
Advanced Topics Guide
SQLBase Optimizer Implementation
Statistics in SYSCOLUMNS
The following statistical information is maintained in these columns of the
SYSADM.SYSCOLUMNS table in the system catalog:
This is the actual average length of this column across all rows in the table.
This can differ from the defined column length since SQLBase stores all
columns as variable data. A base row page stores “a long data description”
for every non-NULL long column.
This is the actual average length of a LONG VARCHAR column across all
rows of the table. This value is zero if this is not a LONG VARCHAR
Index statistics
The statistics associated with indexes on tables are also maintained in two separate
system catalog tables, as detailed in the following paragraphs. All columns (except
one, the HEIGHT column in the SYSINDEXES table) are maintained statically, and
are populated when UPDATE STATISTICS is run or the index is initially created.
Statistics in SYSINDEXES
The following statistical information is maintained in these columns of the
SYSADM.SYSINDEXES table in the system catalog:
The height of the index tree is the number of nodes that have to be read from
the root to the leaf level inclusive. Also, it is called the depth of the tree. The
statistic is also maintained dynamically in the index’s control page. This field
is null for a hash index.
The total number of nodes in the bottom level (leaves) of the index is the
leafcount. Also, it is the number of pages in the index’s sequence set. This
field is null for a hash index.
The total number of page changes that would occur if the entire table was
read through the index’s sequence set. For a perfectly clustered index, this
column is equal to the number of base data pages in the table. At the other
extreme, the highest value for a totally unclustered index is equal to the
number of rows in the table. This field is null for a hash index.
Advanced Topics Guide
Database statistics
The number of primary pages that have been allocated in the base table for
the subject table of a hashed index. It is the same as the number of hash slots
available for distribution of the table’s rows (see Chapter 10, Hashed Indexes,
for more information on the significance of this figure). This field is null for
a B+Tree index.
This is the number of overflow pages that SQLBase allocates for the subject
table of a hashed index. When the table and index are initially created, this
number reflects the number of overflow pages that SQL Base pre-allocates.
After that, this number increases as additional overflow pages are required to
handle hashing collisions. This field is null for a B+Tree index.
For B+Tree indexes, this is the average length of the key across all entries in
the index. This number is needed because SQLBase maintains all index
entries as variable length, minimum prefix only fields (see Chapter 9, B-Tree
Indexes, for more information on these index entries). This field is null for a
hash index.
Statistics in SYSKEYS
The following statistical information is maintained in this column of the
SYSADM.SYSKEYS table in the system catalog:
This is the number of distinct keys in the index for that portion of the key from
the first column up to the COLNO of this entry. If every possible combination
of values for a composite key exists, this number is the product of each
successive column’s cardinality. Cardinality is the number of distinct values
that a column has in a table. The simplest example is a column called
SEX_CODE, which has a cardinality of two, for the values ‘M’ and ‘F’.
Consider the following index:
The employee table contains 250 rows, one for every employee in the
company. Also, the cardinality characteristics of the columns are as follows:
DEPT - There are 10 departments.
Advanced Topics Guide
SQLBase Optimizer Implementation
SALARY - No two employees in the company have the exact same
salary; in other words, there are 250 discrete salary values.
YEARS_SERVICE - The company experienced three distinct hiring
periods, and no one has ever left. Therefore, there are only three distinct
values for this column.
Given these conditions, the SysKeys rows which describe these index
columns have the following values for DISTCOUNT:
DEPT - 10
One for each department in the company.
SALARY - 250
Since each of the employees has a discrete salary value, the number of
keys at this point in the index equals the number of rows in the table.
Alternatively, if there are only five salaries paid by the company, and
each department has someone at each of the five salary levels, then the
DISTCOUNT for this column is 50 (10 times 5).
Since the salary field had already raised the cardinality of the index keys
to be equal to the number of rows in the table, this field cannot raise it
any higher. It is impossible for the number of index entries at any level
to exceed the number of table rows. However, in the alternate scenario
where there are only five salary levels, the value of DISTCOUNT at this
column is 150 (3 times 50), assuming that every department had people
at each salary level who were hired in each of the three hiring waves.
Selectivity factor
The selectivity factor is a statistic which is computed by the optimizer for predicates
used by the relational select operator (recall that the select operation is implemented
internally in all SQLBase file access operators). In addition, the selectivity factor is
used by the optimizer to determine if an index access path is more efficient than a
table scan. Consequently, SQLBase only calculates the selectivity factor for those
predicates which contain attributes for which an appropriate index exists.
The selectivity factor estimates the number of rows which are returned following the
application of the select predicates on a table of known size. This is critical when the
optimizer is evaluating an index in an access plan. The index will be more efficient
than a table scan only if the selectivity factor multiplied by the number of rows in the
table are less than the number of database pages occupied by the table.
In other words, if most of the table rows have to be retrieved anyway, using the index
only incurs the overhead of reading through the index structure without saving the
Advanced Topics Guide
Database statistics
need to read the same number of data pages. Since this is the case, the selectivity
factor appears in virtually every I/O costing formula associated with an index access
operation that is used by the optimizer.
Numerically, the selectivity factor is a probability, which means it ranges from zero to
one. Multiplying the number of rows in a table by the selectivity factor associated
with that table’s predicates will directly yield the number of rows which would
survive the select operation, based on the assumption that the values of the table are
evenly distributed among its rows.
Single predicates
Determining the selectivity factor of a single predicate, such as EMPLOYEE_NO =
65, depends on what operator is used in the predicate. The operator affects the
selectivity factor because it determines the relationship between the rows that satisfy
the predicate and the other operand.
The more restrictive the operator, the lower the selectivity factor of the predicate. The
most restrictive operator is equality (=), since only a single column value can satisfy
the predicate. In this case, the selectivity factor is simply the inverse of the cardinality
of the column, which is stored in the DISTINCTCOUNT column of the
SYSADM.SYSINDEXES table entry for this index’s key.
However, it is not as easy to calculate the selectivity factor for inequality operators.
Consider the following two predicates:
ORDER_DATE > JAN-01-1900
Obviously, the first predicate could be expected to return many more rows than the
second when applied to the ORDER table. But how can the SQLBase optimizer
determine this? It turns out that the optimizer calculates the selectivity factor by
accessing the upper levels of the index structure which contains the column in
question (in this example, an index on ORDER_DATE). It then estimates the number
of rows that satisfies the predicate by extrapolating the proportion of index keys at this
index level. This technique, called the BTree scan, allows the SQLBase optimizer to
make reasonably accurate estimations about the results of inequality predicates.
The other factor that plays a role in determining the selectivity factor of a single
inequality predicate is whether the column is compared to a constant or a bind
variable. This is not critical for equality predicates because the optimizer can
calculate the selectivity factor from the DISTINCTCOUNT column with the
assumption of equal value distribution throughout the table. This assumption allows
the optimizer to be insensitive to the actual value of the operand on the other side of
the equality operator. However, for the optimizer to use the BTree scan technique, it
must be able to find a value for the other operand. When the operand is a bind
variable, the optimizer must fall back to making an assumption that is actually a hardcoded value.
Advanced Topics Guide
SQLBase Optimizer Implementation
The following table summarizes the determination of the selectivity factor of a single
predicate, depending on the operator and whether a constant or bind variable is used
as the other operand. Note that where ‘transformation’ is listed, the predicate is
eliminated from the query through semantic transformation to another form.
Predicate Type
Operator Constant
Operator Not Constant
!=, <>
1 - 1/cardinality
btree scan
btree scan
btree scan
btree scan
btree scan
btree scan
btree scan
treated as a zero length constant
not applicable
btree scan
Multiple predicates
When multiple predicates are placed on a single table, the SQLBase optimizer
follows some fixed rules when combining their selectivity factors into a total
selectivity factor for the entire relational select operation. The first rule regards how
many predicates are considered when calculating a total selectivity factor for
predicates containing a mix of equality and inequality comparison operators.
SQLBase follows these rules in determining which predicates to evaluate a selectivity
factor for:
When the high order columns of the index key are contained within equality
predicates, the optimizer can find the selectivity factor for the multiple
columns by reading the DISTINCTCOUNT column of the lowest order key’s
SYSKEYS row. When the first inequality predicate is encountered, it is
ignored along with any remaining predicates containing columns that appear
in the index. An example of this is:
Advanced Topics Guide
Use of statistics by the optimizer
SALARY > 25000
Assume the index is defined on a composite key consisting of
EMPLOYEE_NO, then DEPT, followed by SALARY. The combined
selectivity factor for both the EMPLOYEE_NO and DEPT columns, which
appears in the DISTINCTCOUNT column of the DEPT row of the
SYSKEYS catalog table, is used by the optimizer to estimate the number of
rows that will be returned from the query. The effect of the inequality
predicate on the SALARY column is ignored. For an example of how the
combined DISTINCTCOUNT is calculated for multiple columns, see the
database statistics section earlier in this chapter.
If the high order column of the index key is the subject of an inequality
predicate, only the selectivity factor of this predicate is evaluated, and any
remaining predicates are ignored (regardless of whether they are inequality or
equality predicates). An example of this is:
SALARY > 25000 AND
Assume the index is defined on a composite key consisting of SALARY, then
DEPT, followed by YEARS_SERVICE. The optimizer determines the
selectivity factor by looking at the root page of the index to determine what
proportion of table rows have a SALARY > 25000. This proportion becomes
the selectivity factor for the entire group of predicates. The DEPT and
YEARS_SERVICE predicates are ignored.
Use of statistics by the optimizer
The first step the SQLBase query optimizer performs in preparing a statement for
execution is to build a list of candidate access paths that could be executed to satisfy
the data requirements of the SQL statement. Because the optimizer is cost-based,
database statistics are used as the basis for projecting an estimate of the work required
to execute each possible access plan considered by the optimizer as a candidate for
possible execution.
After the SQLBase query optimizer builds a group of alternative access plans, each
plan is assigned an estimated execution cost so that the most inexpensive plan may be
chosen for subsequent execution. The two components of query execution cost
evaluated by SQLBase are I/O and CPU cost. These costs are then combined in order
to assign the plan an overall cost. Once all alternative plans are assigned a cost, the
smallest is selected and the remainder are discarded.
Advanced Topics Guide
Advanced Topics Guide
Chapter 18
Working with the
SQLBase Query
In this chapter we:
Discuss the specific actions that may be taken in order to improve the
performance of a query which is causing its transaction to fail to meet
response time objectives.
Specify a seven step procedure that may be applied to tune the performance
of a query in a systematic manner.
Advanced Topics Guide
Working with the SQLBase Query Optimizer
Section 1, Database Design, presents a methodology for physical database design.
This database design method compares general principles and heuristics to create an
internal schema that exhibits good performance characteristics for most SQL queries.
However, the physical design method also acknowledges that some individual queries
may not meet their performance requirements without further tuning (see chapter 7,
Physical Design Control).
A myriad of tuning options faces database designers for increasing the performance
of individual queries. One of the primary ingredients for success is knowing which
option to use in which situation and when. However, one primary option is working
with the SQLBase query optimizer to ensure that the correct access paths are
available to the query optimizer. This chapter addresses this issue.
When working with the SQLBase optimizer to increase the performance of a specific
SQL SELECT statement, the database designer performs three distinct activities:
update statistics, determine the optimal index set, and override the query plan(s)
chosen by the SQLBase optimizer.
Update statistics
As described earlier in this section, SQLBase maintains several statistics regarding
tables and indexes. The query optimizer uses these statistics to calculate the cost of
various access paths. Therefore, the first aspect of working with the query optimizer,
improving the performance of a given query, is to make sure these statistics are
correct and up-to-date.
Optimal index set
Another aspect of working with the query optimizer involves adding indexes to
increase the speed of restrictions and joins, and to eliminate sorts for GROUP BY and
ORDER BY clauses. If a SELECT command is performing poorly, it is probably
either joining, sorting, or reading large tables. Indexes can increase the performance
of all these operations.
As discussed in chapter 6, a common indexing guideline is to create indexes on all
primary and foreign keys. This is because in most systems these are the columns that
appear most frequently in WHERE clause predicates and ORDER BY and GROUP
BY clauses. This initial index set (the collection of all indexes in a database) provides
indexes for restrictions and eliminates sorts across primary and foreign keys.
Furthermore, joins are frequently the most time-consuming aspect of a specific query.
The fastest join algorithm for large tables is the index merge. The initial index set
ensures that the index merge access path is available to the optimizer as long as the
join columns are either primary or foreign keys. For all other joins, the optimizer is
Advanced Topics Guide
Tuning a SELECT statement
limited to the slower join algorithms, the hash join or one of the nested loop
techniques. However, for most SELECT commands, the access paths available to the
optimizer with the initial index set provide adequate performance.
On the other hand, the initial index set may not meet the performance requirements of
all queries. For example, some queries containing join columns that are neither
primary nor foreign keys may perform poorly. For those queries not meeting their
performance criteria, you must add additional indexes to increase the speed of the
join. However, you do not want to create one or more specific indexes for every
SELECT command performing poorly since:
Indexes slow update DML commands, specifically all INSERT and DELETE
statements, as well as any UPDATE statements that modify a column that is
contained within one or more indexes.
Index management is both time-consuming and costly.
Indexes take additional disk space.
Indexes can also cause locking contention because many rows may be stored
on a single index node page.
Consequently, the general design goal is to create an index set that contains the fewest
possible indexes which meets the performance requirements of all database
transactions. This is the optimal index set.
Tuning a SELECT statement
This section presents a procedure for working with the SQLBase optimizer to tune a
query that is performing poorly. The procedure iterates to the optimal index set. We
also assume you are familiar with the protocol described in chapter 7 for physical
design control. The procedure consists of seven discrete steps, each of which are
described in the following paragraphs.
Step 1: Update statistics
Before adding indexes, verify that the database statistics stored in the system tables
are correct. If you execute the query against an actual production database, you should
update the statistics for all the tables listed in the FROM clause using the UPDATE
STATISTICS command. On the other hand, if you are using a small test database, you
may have to manually calculate and override the statistics calculated by SQLBase for
both the tables listed in the FROM clause and their indexes.
Once you have updated the statistics, you should compile the SQL command with the
PLANONLY parameter set ON. Compare the new query plan to the old plan before
updating the statistics to determine if the plan has changed. (Sometimes it requires
several minutes or hours to execute a query. By comparing the plans, you can avoid
executing a query only to find that the performance is identical to the last time the
Advanced Topics Guide
Working with the SQLBase Query Optimizer
query was executed). If so, execute the query to determine if performance has
improved significantly.
Step 2: Simplify the SELECT command
Before adding indexes or forcing plans, you should simplify the query. The objective
is to make the SQL for the SELECT statement as simple as possible in order to
eliminate as many variables as possible. Once you have simplified the query, compile
the SQL command with the PLANONLY parameter set ON. Compare the new query
plan to the plan before simplifying the command to determine if the plan has
changed. If so, execute the query to determine if performance has improved
To simplify the SELECT command:
Eliminate unnecessary predicates or clauses
Add parentheses to arithmetic and Boolean expressions
Convert bind variables to constants
Eliminate unnecessary predicates or clauses
It is common to include more SQL clauses in a SELECT command than is absolutely
necessary. This is usually done to ensure the correct answer is returned and stems
from a lack of confidence in one’s understanding of the SQL syntax. Before
attempting to tune the query, the database designer should remove all unnecessary
clauses and predicates. Clauses which are commonly included but not needed are:
ORDER BY clauses. Frequently, an ORDER BY clause is included even
though an exact ordering of the result set is not required by the application
program or end-user.
WHERE clause predicates. Frequently, WHERE clauses contain
redundant restriction predicates. For example, the predicates in the following
WHERE clause are redundant, since DEPT_NO is the primary key, and
therefore will uniquely identify only one row:
Add parentheses to arithmetic and Boolean expressions
The SQLBase SQL language syntax includes a specific precedence for evaluating
arithmetic and Boolean expressions. However, it is easy to make a mistake in
applying the precedence rules. Therefore, we recommend you add parentheses to
these expressions to explicitly state the precedence. You may think the optimizer is
evaluating an expression in one manner when it is actually using another.
Advanced Topics Guide
Tuning a SELECT statement
Convert bind variables to constants
Bind variables are general place holders used when compiling an SQL command. It is
often desirable to use stored commands to eliminate the overhead of compilation
when one SQL statement will be executed many times. These stored commands are
pre-compiled, and their access plan is saved in the system catalog along with their
source text. Also, programs sometimes issue one prepare for a statement that will be
executed many times throughout the program with little modification. In these
situations, bind variables are used to allow the values of some variable tokens in the
command to change between executions without having to re-compile. However,
when bind variables are used, the query optimizer does not know what values are
bound to the program variables when the SQL command is executed. This causes the
query optimizer to make some gross assumptions when calculating the selectivity
factor. These assumptions may lead to a high selectivity factor which can prevent the
optimizer from using a specific index. Consequently, when tuning queries, you should
convert all bind variables to constants so that the query optimizer has specific values
to work with.
For example, when bind variables are used in restriction predicates, the query
optimizer is forced to use a default selectivity factor of 1/3. Consider the following
Amount > :BindVariable
Amount > 1000
The advantage of the first predicate is that the SQL command can be compiled once
and executed many times using many values in the bind variable. The disadvantage is
the query optimizer has less information available to it at compile time. The optimizer
does not know whether 10 or 10,000,000 will be bound to the program variable.
Hence, the optimizer is unable to calculate the selectivity factor. Consequently, in
these situations the optimizer simply sets the selectivity factor to a constant that may
not reflect the true situation (this constant is currently 1/3 for SQLBase 6.0, but is
subject to change in future releases). Since this 1/3 figure is relatively high (in the
absence of other more restrictive predicates) the optimizer may not use any index
associated with the column contained in these predicates.
On the other hand, with the second predicate, the optimizer knows the comparison
value at the time of the compile and is therefore able to calculate a more accurate
selectivity factor.
Bind variables also cause problems when used in LIKE predicates. The LIKE
operator can include wild card characters. For example, the following query finds all
customers whose name begins with the letter A:
select * from customer where name like ‘A%’
The wild card characters can occur in the beginning (for example, '%A'), middle (for
example, 'A%A'), and terminal positions. The nature of BTree indexes is such that
Advanced Topics Guide
Working with the SQLBase Query Optimizer
these indexes are only helpful in queries involving the LIKE operator when the wild
card characters do not occur in the beginning. Therefore, the SQLBase optimizer
does not use an index if a LIKE operator is used and a wild card is in the beginning
Furthermore, the SQLBase optimizer also does not use an index when bind variables
are used with the LIKE operator. This is because the optimizer does not know at
runtime where the wild card will occur. Therefore, the optimizer chooses an access
path (i.e., table scan) which satisfies the query with the wild card in any position,
resulting in significantly decreased performance. Hence, convert all bind variables to
constants in LIKE predicates.
Step 3: Review the query plan
Execute the query with the PLANONLY parameter ON to display the query plan
which the optimizer has chosen. Familiarize yourself with the query plan since you
must understand the plan in order to use it. Several items to especially look for are:
Conversion of subqueries to joins: The optimizer converts most
subqueries to joins. Make sure you understand where these transformations
are taking place; you need to know where these transformations occur in
some of the following steps.
Temporary tables being created: If a temporary table is created, this can
indicate that the optimizer is sorting intermediate results. If this is occurring,
you may want to add an index in one of the following steps to avoid the sort.
Slower join techniques: The hash join and nested loop techniques are not
as fast as the index merge technique for large tables. If these techniques are
being used, you may want to add indexes in steps 5 or 6 so that these joins
use the index merge technique. Of course, for some situations the hash join
may represent the best possible method when large amounts of data have to
be processed.
Step 4: Locate the bottleneck
The query which is performing poorly can contain many clauses and predicates. If so,
you must determine which clauses or predicates are resulting in the poor
performance. Generally if you remove one or two clauses or predicates, query
performance improves significantly. These clauses are the “bottleneck.”
Experiment with the query; iteratively remove (comment out) clauses and/or
predicates until the query exhibits substantially superior performance. Locate the
bottleneck clauses and predicates. For example:
Advanced Topics Guide
Tuning a SELECT statement
If the query contains an ORDER BY clause, comment it out and see if the
query plan changes. If the query plan changes, execute the query to determine
if performance improves significantly.
If the query contains several joins, locate the one that is taking the longest to
perform. Successively comment out all but one join and execute the query.
Determine if one join is the bottleneck.
Eliminate any AT (@) functions (such as @UPPER) being performed on
WHERE clause predicates and see if performance increases. It may be that an
index is not being used because of a @function. If this is the case, the index
may be altered to contain the function if the query and function are
sufficiently critical. Of course, this alternative may then make the index
unusable for other queries whose performance had been acceptable before the
Step 5: Create single column indexes for the bottlenecks
If you complete steps 1 through 4 and the query is still not meeting its performance
requirements, you must create indexes specifically for the query to increase
performance. In general, single column indexes are preferable to multi-column
indexes because they are more likely to be useful to other queries. Try to increase the
performance of the query with single column indexes first before resorting to multicolumn indexes.
Locate the following table columns used in the query which do not have an index
(either a single attribute index or a right-hand subset of a multi-column index):
Join columns : This includes subqueries that the optimizer converts to
joins. If the primary key of the table is a composite key, the join is specified
across multiple columns and a multi-column index is needed to increase the
performance of this join. This index should already exist.
GROUP BY columns: If the GROUP BY clause contains more than one
column, a multi-column index is needed to increase the performance of this
GROUP BY clause. Skip these columns until step 6.
ORDER BY columns: If the ORDER BY clause contains more than one
column, a multi-column index is needed to increase the performance of this
ORDER BY clause. Skip these columns until step 6.
Low cost restrictions: These are the table columns referenced in the
WHERE clause restriction predicate with the lowest selectivity factor for
each table reference in the FROM clause.
Compare these table columns to the bottlenecks discovered in step 4. One of the table
columns identified above should match one of the bottleneck clauses or predicates
identified in step 4. If so, create a single column index on one of these columns.
Determine if the addition of this index changes the query plan by running the query
Advanced Topics Guide
Working with the SQLBase Query Optimizer
with the PLANONLY parameter ON. If it does, execute the query to determine if
adding this index significantly increases the performance of the query.
If adding the index does not increase performance significantly, create another index
on another table column identified above. Continue until you have tried all columns
identified above. If you are unable to increase the performance of the query by adding
a single column index, proceed to the next step.
Step 6: Create multi-column indexes
As a last resort, you may have to create one or more multi-column indexes
specifically for this query to achieve the desired performance. The procedure for
doing so is identical to the one described in the previous section for single column
Create a multi-column index.
Determine if the query plan changes. If it does, execute the query to determine if
performance increases.
If performance does not improve, create another index. Continue creating indexes
until all possible multi-column indexes have been created.
The types of multi-column indexes to create are listed below:
Multi-column GROUP BY clause: If the query has a GROUP BY clause
containing more than one column, create an index on all of the columns listed
in the GROUP BY clause. Be sure to specify the columns in the same order
as listed in the GROUP BY clause.
Multi-column ORDER BY clause: If the query has an ORDER BY clause
containing more than one column, create an index on all of the columns listed
in the ORDER BY clause. Be sure to specify the columns in the same order
as listed in the ORDER BY clause. Also, be sure to specify the same
sequencing (ascending or descending) for each column as specified in the
ORDER BY clause.
Join column(s) plus the low cost restriction : For each table listed in
the FROM clause that has at least one restriction specified in the WHERE
clause (in descending order from the largest table to the smallest), create a
multi-column index on the join column(s) and the column in the restriction
predicate with the lowest selectivity factor.
Join column(s) plus all restrictions: For each table listed in the FROM
clause that has at least one restriction specified in the WHERE clause (in
descending order from the largest table to the smallest), create a multicolumn index on the join column(s) and all columns included in any
restrictions involving that table. The order of the columns in the index is
critical to whether the index will be selected for use by the optimizer. In
Advanced Topics Guide
Tuning a SELECT statement
general, list equality columns before other columns, and list all equality
predicate columns in increasing selectivity factor order. After listing all
equality predicates in the CREATE INDEX statement, list the inequality
predicate which has the lowest selectivity factor as the last column in the
index. Any other inequality predicates present in the WHERE clause will
probably not benefit the performance of the statement, nor will they be
considered by the SQLBase query optimizer when selecting access paths.
Step 7: Drop all indexes not used by the query plan
As discussed in the previous section, indexes slow down UPDATE DML commands
and index management is time-consuming and costly. Therefore, you should keep
track of the indexes created during steps 5 and 6 and drop all indexes that are not used
by the query.
Advanced Topics Guide
Advanced Topics Guide
abbreviations 2-3
access path selection 8-2
access plans 16-2
matching index scan 16-10
non-matching index scan 16-10
physical operators 16-2, 16-9
relational join 16-11
relational operations 16-4
aggregate query 12-6
data 7-11
purpose 16-9
scalar 16-9
ALTER TABLE 3-9, 6-6
conceptual schema 1-2, 2-1, 4-2–4-3, 7-2
external schema 1-2, 1-4, 1-5, 4-4–4-5, 5-7, 7-2
views as 4-2, 4-4
internal schema 1-2, 1-5, 1-6, 3-2, 3-5, 5-7, 6-1,
three schema architecture 1-2–1-6, 4-2, 6-1, 7-2
aggregate data 7-9
derived data 7-9
authorization-id 3-3
index-only access 16-14
index-only access with project operator 16-14
sequence set 16-10
B+Tree index, compared to hashed index 10-2
Bachman diagram 2-6, 2-7
binary search tree
balanced 9-4, 9-6
branching factor 9-3
key value 9-3
node key 9-3
pointer 9-3
short version for indexes 9-3
target key 9-3, 9-4, 9-8
unbalanced 9-6
bind variables 18-5
blocking factor 6-4
Boolean operators 16-5
brute force method 7-3–7-4
B-tree 18-5
general purpose indexing 9-2
index structures 9-3
leaf node 9-4, 9-5, 9-6, 9-7, 9-8, 9-9, 9-12, 9-13
root page 17-9
scan 17-7
bucket size 10-7
cache 14-11–14-13
dirty cache buffers 14-5
dirty cache pages 14-11
used in join operation 16-12
Cache Manager 14-11
cardinality 2-7–2-10, 7-12, 17-5
column 6-11
low 6-16
maximum 2-7, 2-8, 3-6
minimum 2-7, 2-8
Cartesian product 15-4, 16-4
CASE tool 3-2, 5-3
CHAR 3-4, 6-4
Chen diagram 2-6, 2-10
child table 3-2, 3-5, 3-8, 3-9, 6-6, 6-9
clustered hashed index 10-2
columns, adding 3-3
conjunctive form of multiple predicates 16-17
conventions xvii
conversion of subqueries 18-6
correlation variables 16-6, 16-8
packing density 10-13
CREATE INDEX 6-9, 6-14
CREATE TABLE 6-9, 6-13, 6-18
CREATE UNIQUE INDEX 3-9, 6-9, 6-13–6-14
cursor context preservation(CCP)
effect on result set flushing 12-8
rules for releasing locks 13-8
Cursor Stability (CS), drawback 13-11
data administration 2-2, 2-4, 2-5
Advanced Topics Guide
data control language (DCL) 4-4
data definition language (DDL) 3-2, 3-9
data dictionary 2-4
data element 2-3
data integrity 5-2
data pages
page 8-2
page buffers 7-4
page size 6-3
data pages, SQLBase types
extent page 8-3, 8-4
overflow 8-3, 8-4
row pages 8-3, 8-4
data typing columns 3-3
administration 2-2
design based on subject 2-5
locking example 13-4
partitioning 6-16
standardizing names 2-4
transaction 7-6
database manipulation language (DML) 7-5
limitations 4-2
DATE 3-4
deadlock avoidance 13-11
decision support systems 5-4
decoupling programs from data, benefits of 15-3
denormalization 2-18, 7-4, 7-7–7-9, 7-11
techniques 6-2
dependency on data structure 15-2
dependency rule 3-5
diagrammatic symbols 2-6
dimension 2-11
direct random access 6-7
disk access operations 16-10
disk controller cache 7-4
DML, see data manipulation language
domains 2-6, 2-11, 3-3
domain specification 3-2
equality predicate 6-12
equijoin 16-7
explain plan facility 7-5
exponent-mantissa format 3-4
extent page 8-7
extents 10-11
E.F. Codd 2-11, 16-3
enterprise-wide data model 2-4, 2-5
entities, major 2-6
entity relationship diagram (ERD) 2-5, 2-6, 3-2–3-5,
3-6, 3-8, 5-2, 7-2
Advanced Topics Guide
FETCHTHROUGH 12-12–12-13
buffer 8-2
file names 2-3
method 7-3, 7-4
model 7-4–7-5
foreign key column 16-7
free space 6-10, 6-17
calculating 6-18
FROM 16-6
fuzzy checkpointing 14-12–14-13
GROUP BY 16-8, 16-9
access to disk 16-11
hashed table 7-7
hash algorithm, eliminating bias 10-3, 10-4
example 10-4
hash bucket calculation 10-14
hash function 6-7
hash function, SQLBase pre-programmed 10-11
hash key selection 10-6
hashed table 16-11
hashed tables
when not to use 10-3
hashing performance analysis 10-8
history file 13-13
incremental lock(U-Lock) 13-5
B+ trees 6-7
candidates 6-12
good 6-14
poor 6-12, 6-15–6-16
clustered hashed 6-4, 6-7–6-9, 6-10, 7-7
composite index 6-12, 6-16
index only access 6-12–6-13
key 6-11, 6-15
index locking 13-8
index set 18-2
index, effect on performance 13-9
Information Engineering diagram 2-6
inner table 16-11
INSERT 6-6, 6-10
INSERT, locks used in process 13-8
insertion algorithm 9-7
intermediate nodes 9-5
intersection 16-3
isolation mode
changing during transaction 13-10
criteria for choosing 13-14
design trade-offs 13-13
isolation mode, effect on lock duration 12-10
join 3-4, 4-2, 16-4, 16-6–16-8
examples 16-6
hash 16-13
index merge 16-13
loop join with index 16-12
loop with hash index 16-12
merge join 15-6
nested loop join 15-5
outer table 16-11
outerjoin 16-7
selfjoin 16-8
semijoin 16-7
simple loop join 16-12
techniques 18-6
data 2-10, 3-8
table 3-2, 3-7, 3-8
composite 2-11, 3-7
concatenated 2-11, 2-12
foreign 2-11, 2-13, 3-2, 3-5, 3-8, 6-5–6-6, 6-8–610, 6-14, 7-8
clustering 6-8
primary 2-5, 2-11, 2-11–2-13, 3-2, 3-5, 3-7–3-9,
6-2–6-6, 6-7, 6-9–6-11, 6-13–6-15
primary index 6-6
key columns 16-7
skewness of 6-15
LIKE 18-5
linked lists 8-3
active 14-5
archive 14-5
LOGBACKUP 14-5–14-7
manager 14-7
pinned 14-6
log file 13-3
logical data modeling 2-6, 2-13, 5-2
logical data structures 5-2
logical database design process 1-5
logical file
address 8-2
offset 8-2
lost updates, causes 13-5
metadata 2-4
multiversion concurrency control 13-13
naming standards 2-2, 2-4
non-participating rows 16-7
normal form 2-14, 2-18
normalization 2-6, 2-12, 2-13, 4-2
notation conventions xvii
null usage 3-4
conceptual 4-2
owners 4-2
physical 4-2
online transaction processing 5-3
Advanced Topics Guide
optimizer 6-7, 6-12, 7-3, 7-5–7-6, 7-13
cost-based 16-2
overflow page 10-11
execution cost 17-9
query plan, see access plans
query tree 16-15–16-16
packing density 10-7
70% rule-of-thumb 10-8
page clustering 6-4
page number
logical 8-4
physical 8-2, 8-4, 8-7, 8-11
page size
DOS,Windows 11-3
SQLBase 8-2, 8-7
page types
control 8-3
data 8-3
index 8-3
parent 6-6, 6-9, 6-14
parent table 3-5, 3-6–3-8
partition and merge algorithm 16-9
partnership set 2-8
PCTFREE 6-18–6-19
performance requirements tailoring 6-5
physical block reads 7-4
physical block writes 7-4
physical database
dependencies 4-2
design 5-2, 5-4–5-6, 7-2
schema 4-2
physical design control 7-2–7-4, 7-6
physical schema 7-3, 7-6
predicate 16-5
conjunctive, disjunctive 16-5
multiple 17-8
short-circuit testing 16-14, 16-18
primary key 4-3
primary key index 4-3
primary table 7-8
prime attributes 7-11
printer names 2-4
probe phase 16-13
program names 2-3
projection 16-4, 16-5, 16-14
random reads 8-2
read locks(S-Locks) 13-5
Read Only (RO) 13-12
REAL 3-4
Recovery Manager 14-2
referential integrity 2-11, 3-2, 3-8–3-10, 4-3, 6-5, 614–6-15
CASCADE 3-9, 6-6
constraints 6-5–6-6
SET NULL 3-5, 3-9
regular node 9-5
relational algebra 16-3
relational view 4-2
relationship 2-6
information bearing 2-10
many-to-many 3-2, 3-6
one-to-many 2-13, 3-6–3-8
one-to-one 2-13, 5-3, 6-5
resolving 2-10, 3-6
Release Locks(RS), example 13-12
reorganization frequency 6-18
Repeatable Read (RR) 13-11
repository 2-4
result sets
derived 12-7
implications for transactions 12-9
ROLLBACK 13-2, 14-2, 14-6
row clustering 7-6
row serial number 8-6–8-7, 8-18
row size 6-3–6-4
ROWID 12-2–12-7, 12-11–12-12
rowid 8-7
rows per page 6-4
Advanced Topics Guide
secondary table 7-8
security administration 4-3–4-4
security authorizations 2-3
SELECT 6-7, 6-11–6-13, 6-15, 16-4, 18-4
criteria 7-7, 7-11
description 16-14
examples 16-5
selectivity factor 6-11–6-12, 6-16, 7-12, 16-14, 17-6,
for inequality operators 17-7
total 17-8
semantic validation 2-11
sequencing node properties 9-5
sequential reads 8-2
server names 2-4
set operations 16-3
byte offset 8-6
row header 8-5
table 8-4–8-9
sort key 16-9
splitting tables 6-2–6-6
horizontal splitting 6-2, 6-4
vertical splitting 6-2
SQLBase catalog 16-2
sqlexp function 7-6
SQLPPCX flag 13-8
standardized templates 2-3
static data 7-7
UPDATE STATISTICS 7-6, 17-2, 17-4, 18-3
user modifiable 17-2
strategic data plan 2-4, 2-6
strategic management 2-4
system catalog tables 17-2
system development 2-2
table names 2-3
table scan 8-2, 16-10
table size calculation
hashed 11-6
regular 11-6
temporary table 12-6, 18-6
TIME 3-4
batch 5-4
complexity 5-4, 5-5
data manipulation language (DML) 1-4
database 5-2
defining 5-2, 5-3
definitions 5-2–5-6
description 5-3
high complexity 5-4, 5-5
low complexity 5-4, 5-5
medium complexity 5-5
name 5-3
number 5-3
online 5-4
performance 5-2, 6-4
performance requirements 5-5
relative priority 5-5
secondary effects example 13-4
SQL statements 5-6
type 5-4
volumes 5-4
transformation rules 15-5
multiway 9-2, 9-4
tuning the schema 7-6
uncommitted data
dependency problem 13-5
effect on other transactions 13-3
union 16-3
update serial number 8-5–8-7
user names 2-3
Advanced Topics Guide
VARCHAR 3-4, 6-4–6-5
view administration 1-5
view names 2-3
virtual table 12-7
WHERE 6-7, 6-8, 6-12–6-13, 16-6
workstation names 2-4
write locks(X-Locks) 13-5
Advanced Topics Guide
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF