What Is NoSQL?
SYSADMIN
NoSQL
GREG BURD
Greg Burd is a Developer
Advocate for Basho
Technologies, makers of Riak.
Before Basho, Greg spent
nearly ten years as the product manager for
Berkeley DB at Sleepycat Software and then
at Oracle. Previously, Greg worked for NeXT
Computer, Sun Microsystems, and KnowNow.
Greg has long been an avid supporter of open
source software.
[email protected]
Choosing between databases used to boil down to examining the differences
between the available commercial and open source relational databases. The term
“database” had become synonymous with SQL, and for a while not much else came
close to being a viable solution for data storage. But recently there has been a shift
in the database landscape. When considering options for data storage, there is a
new game in town: NoSQL databases. In this article I’ll introduce this new category of databases, examine where they came from and what they are good for, and
help you understand whether you, too, should be considering a NoSQL solution in
place of, or in addition to, your RDBMS database.
What Is NoSQL?
The only thing that all NoSQL solutions providers generally agree on is that the
term “NoSQL” isn’t perfect, but it is catchy. Most agree that the “no” stands for “not
only”—an admission that the goal is not to reject SQL but, rather, to compensate for
the technical limitations shared by the majority of relational database implementations. In fact, NoSQL is more a rejection of a particular software and hardware
architecture for databases than of any single technology, language, or product.
Relational databases evolved in a different era with different technological constraints, leading to a design that was optimal for the typical deployment prevalent
at that time. But times have changed, and that once successful design is now a limitation. You might hear conversations suggesting that a better term for this category
is NoRDBMS or half a dozen other labels, but the critical thing to remember is that
NoSQL solutions started off with a different set of goals and evolved in a different
environment, and so they are operationally different and, arguably, provide bettersuited solutions for many of today’s data storage problems.
Why NoSQL?
NoSQL databases first started out as in-house solutions to real problems in
companies such as Amazon Dynamo [1], Google BigTable [2], LinkedIn Voldemort
[3], Twitter FlockDB [4], Facebook Cassandra [5], Yahoo! PNUTS [6], and others.
These companies didn’t start off by rejecting SQL and relational technologies;
they tried them and found that they didn’t meet their requirements. In particular,
these companies faced three primary issues: unprecedented transaction volumes,
expectations of low-latency access to massive datasets, and nearly perfect service
availability while operating in an unreliable environment. Initially, companies
tried the traditional approach: they added more hardware or upgraded to faster
;login: O C TO B E R 2 0 1 1 5
hardware as it became available. When that didn’t work, they tried to scale existing relational solutions by simplifying their database schema, de-normalizing the
schema, relaxing durability and referential integrity, introducing various query
caching layers, separating read-only from write-dedicated replicas, and, finally,
data partitioning in an attempt to address these new requirements. Although each
of these techniques extended the functionality of existing relational technologies,
none fundamentally addressed the core limitations, and they all introduced additional overhead and technical tradeoffs. In other words, these were good band-aids
but not cures.
A major influence on the eventual design of NoSQL databases came from a dramatic shift in IT operations. When the majority of relational database technology
was designed, the predominant model for hardware deployments involved buying
large servers attached to dedicated storage area networks (SANs). Databases were
designed with this model in mind: They expected there to be a single machine
with the responsibility of managing the consistent state of the database on that
system’s connected storage. In other words, databases managed local data in files
and provided as much concurrent access as possible given the machine’s hardware
limitations. Replication of data to scale concurrent access across multiple systems was generally unnecessary, as most systems met design goals with a single
server and reliability goals with a hot stand-by ready to take over query processing in the event of master failure. Beyond simple failover replication, there were
only a few options, and they were all predicated on this same notion of completely
consistent centralized data management. Technologies such as two-phase commit
and products such as Oracle’s RAC were available, but they were hard to manage,
very expensive, and scaled to only a handful of machines. Other solutions available included logical SQL statement-level replication, single-master multi-replica
log-based replication, and other home-grown approaches, all of which have serious
limitations and generally introduce a lot of administrative and technical overhead.
In the end, it was the common architecture and design assumptions underlying
most relational databases that failed to address the scalability, latency, and availability requirements of many of the largest sites during the massive growth of the
Internet.
Given that databases were centralized and generally running on an organization’s
most expensive hardware containing its most precious information, it made sense
to create an organizational structure that required at least a 1:1 ratio of database
administrators to database systems to protect and nurture that investment. This,
too, was not easy to scale, was costly, and could slow innovation.
A growing number of companies were still hitting the scalability and performance
wall even when using the best practices and the most advanced technologies of
the time. Database architects had sacrificed many of the most central aspects of
a relational database, such as joins and fully consistent data, while introducing
many complex and fragile pieces into the operations puzzle. Schema devolved from
many interrelated fully expressed tables to something much more like a simple
key/value look-up. Deployments of expensive servers were not able to keep up with
demand. At this point these companies had taken relational databases so far outside their intended use cases that it was no wonder that they were unable to meet
performance requirements. It quickly became clear to them that they could do
much better by building something in-house that was tailored to their particular
workloads. These in-house custom solutions are the inspiration behind the many
NoSQL products we now see on the market.
6 ;login: VOL. 3 6, NO. 5
NoSQL’s Foundations
Companies needed a solution that would scale, be resilient, and be operationally
efficient. They had been able to scale the Web (HTTP) and dynamic content generation and business logic layers (Application Servers), but the database continued
to be the system’s bottleneck. Engineers wanted databases to scale like Web servers—simply add more commodity systems and expect things to speed up at a nearly
linear rate—but to do that they would have to make a number of tradeoffs. Luckily, due to the large number of compromises made when attempting to scale their
existing relational databases, these tradeoffs were not so foreign or distasteful as
they might have been.
Consistency, Availability, Partition Tolerance (CAP)
When evaluating NoSQL or other distributed systems, you’ll inevitably hear about
the “CAP theorem.” In 2000 Eric Brewer proposed the idea that in a distributed
system you can’t continually maintain perfect consistency, availability, and partition tolerance simultaneously. CAP is defined by Wikipedia [7] as:
Consistency: all nodes see the same data at the same time
Availability: a guarantee that every request receives a response about whether it
was successful or failed
Partition tolerance: the system continues to operate despite arbitrary message loss
The theorem states that you cannot simultaneously have all three; you must make
tradeoffs among them. The CAP theorem is sometimes incorrectly described as a
simple design-time decision—“pick any two [when designing a distributed system]”—when in fact the theorem allows for systems to make tradeoffs at run-time
to accommodate different requirements. Too often you will hear something like,
“We trade consistency (C) for AP,” which can be true but is often too broad and
exposes a misunderstanding of the constraints imposed by the CAP theorem.
Look for systems that talk about CAP tradeoffs relative to operations the product
provides rather than relative to the product as a whole.
Relaxing ACID
Anyone familiar with databases will know the acronym ACID, which outlines the
fundamental elements of transactions: atomicity, consistency, isolation, and durability. Together, these qualities define the basics of any transaction. As NoSQL
solutions developed it became clear that in order to deliver scalability it might be
necessary to relax or redefine some of these qualities, in particular consistency
and durability. Complete consistency in a distributed environment requires a great
deal of communication involving locks, which force systems to wait on each other
before proceeding to mutate shared data. Even in cases where multiple systems are
generally not operating on the same piece of data, there is a great deal of overhead
that prevents systems from scaling.
To address this, most NoSQL solutions choose to relax the notion of complete
consistency to something called “eventual consistency.” This allows each system
to make updates to data and learn of other updates made by other systems within
a short period of time, without being totally consistent at all times. As changes are
made, tools such as vector clocks are used to provide enough information to reason
about the ordering of those changes based on an understanding of the causality of
the updates. For the majority of systems, knowing that the latest consistent infor-
;login: O C TO B E R 20 11 NoSQL 7
mation will eventually arrive at all nodes is likely to be enough to satisfy design
requirements.
Another approach is optimistic concurrency control, using techniques such as
multi-version concurrency control (MVCC). Such techniques allow for consistent
reading of data in one transaction with concurrent writing in another transaction
but do not address write conflicts and can introduce more transaction retries when
transactions overlap or are long-running.
Both eventual consistency and MVCC require that programmers think differently about the data they are managing in the application layer. Both introduce
the potential that data read in a transaction may not be entirely up-to-date even
though it may be consistent.
Achieving durability has long been the bottleneck for database systems. It’s easy
to understand why writing to disk slows down the database; disk access times
are orders of magnitude slower than writes to memory. Most database solutions
recognize the potential performance tradeoffs related to durability and offer ways
to tune writes to match application requirements, balancing durability against
speed. Be careful, as this can mean leaving a small window of opportunity where
seemingly committed transactions can be lost under certain failure conditions.
For example, some will consider an update durable when it has been written into
the memory of some number of systems. If those systems were to lose power, that
commit would be lost. Network latencies are much faster than disk latencies, and
by having data stored on more than one system the risk of losing information due
to system failure is much lower. This definition of durability is very different from
what you’ve come to expect from a relational database, so be cautious. For instance,
Redis, MongoDB, HBase, and Riak range from minimally durable to highly durable,
in that order. As with any other database, when evaluating NoSQL solutions, be
sure to know exactly what constitutes durability for that product and how that
impacts your operational requirements.
Data and Access Model
The relational data model with its tables, views, rows, and columns has been very
successful and can be used to model most data problems. By using constraints,
triggers, fine-grained access control, and other features, developers can create
systems that enforce structure and referential integrity and that secure data.
These are all good things, but they come at a price. First, there is no overlap in the
data representation in SQL databases and in programming languages; each access
requires translation to and from the database. Object to relational mapping (ORM)
solutions exist to transparently transform, store, and retrieve object graphs into
relational databases, and although they work well, they introduce overhead into a
process and slow it down further. This is an impedance mismatch that introduces
overhead where it is least needed. Second, managing global constraints in a distributed environment is tricky and involves creating barriers (locks) to coordinate
changes so that these constraints are met. This introduces network overhead and
sometimes can stall progress in the system.
NoSQL solutions have taken a different approach. In fact, NoSQL solutions diverge
quite a bit from one another as well as from the RDBMS norm. There are three
main data representation camps within NoSQL: document, key/value, and graph.
There is still a fairly diverse set of solutions within each of these categories. For
instance, Riak, Redis, and Cassandra are all key/value databases, but with Cassan-
8 ;login: VOL. 3 6, NO. 5
dra you’ll find a slightly more complex concept, based on Google’s BigTable, called
“column families,” which is very different from the more SimpleDB-like “buckets
containing key/value pairs” approach of the other two.
Document-oriented databases store data in formats that are more native to the languages and systems that interact with them. JavaScript Object Notation (JSON)
and its binary encoded equivalent, BSON, are used as a simple dictionary/array
representation of data. MongoDB stores BSON documents and provides a JSONencoded query syntax for document retrieval. For all intents and purposes, JSON
has replaced most of the expected use cases for XML, and although XML has other
advantages (XQuery, typed and verifiable schema), JSON is the de facto Webnative format for data on the wire and now can be stored, indexed, and queried in
some NoSQL databases.
Neo4j is a graph-based NoSQL database that stores information about nodes and
edges and provides simple, highly optimized interfaces to examine the connectedness of any part of the graph. With the massive growth of social networking sites
and the value of understanding relationships between people, ideas, places, etc.,
this highly specialized subset of data management has received a great deal of
attention lately. Look for more competitors in this space as the use cases for social
networks continue to grow.
Global constraints are generally not available or very rudimentary. The reasoning for not introducing more than basic constraints is simple: anything that could
potentially slow or halt the system violates availability, responsiveness, and
latency requirements of these systems. This means that applications using NoSQL
solutions will have to build logic above the database that monitors and maintains
any additional consistency requirements.
A crucial part for the success of the relational market as a whole has been the relative uniformity of SQL solutions. Certainly there are differences between vendors
and a non-trivial penalty when migrating from one SQL solution to another, but at
a high level, they all start with the same APIs (ODBC, JDBC, and SQL), and that
allows a universe of tools and a population of experts to flourish. NoSQL solutions
also diverge when it comes to access, encoding, and interaction with the server.
While some NoSQL products provide a HTTP/REST API, others provide simple
client libraries or network protocols that use widely available data encoding libraries such as Thrift and Protobufs. Some provide multiple methods of access, and the
more successful NoSQL solutions will generally have pre-built integrations with
most of the popular languages, frameworks, and tools, so that this is not as big an
issue as it may seem.
This diversity is representative of the fact that NoSQL is a broad category where
there are no standards such as SQL to unify vendors. Customers should therefore
choose carefully—vendor lock-in is a given at this point in the NoSQL market. That
said, most leading NoSQL solutions are open source, which does defray some of the
risk related to the sponsoring company changing hands or going out of business.
In the end, though, moving from one NoSQL solution to another will be a timeconsuming mistake that you should try at all costs to avoid.
Distributed Data, Distributed Processing
NoSQL solutions are generally designed to manage large amounts of data, more
than you would store on any single system, and so all generally have some notion of
;login: O C TO B E R 20 11 NoSQL 9
partitioning (or sharding) data across the storage found on multiple servers rather
than expecting a centrally connected SAN or networked file system. The benefits
of doing this transparently are scalability and reliability. The additional reliability comes when partitions overlap, keeping redundant copies of the same data at
multiple nodes in the system. Not all NoSQL systems do this. The drawback will
be some amount of duplicated data and costs associated with managing consistency across these partitions. In addition it is critical to understand the product’s
approach to distributing data. Is it a master/replica, master/master, or distributed
peers? Where are the single points of failure? Where are there potential bottlenecks? When reviewing the NoSQL solutions it’s important not to gloss over the
details, and be sure to run extensive in-house tests where you introduce all manner
of failure conditions into your testbed. The good NoSQL solutions will prove resilient even when operating in punishing environments, whereas the less mature may
lose data or stop operating when too many unforeseen conditions arise.
In addition to distributing data, many NoSQL solutions offer some form of distributed processing, generally based on MapReduce. This can be a powerful tool when
used correctly, but again the key when evaluating NoSQL solutions is to dig into
the details and understand exactly what a vendor means by “we support MapReduce.”
NoSQL in Practice
There are many products that now claim to be part of the NoSQL database market,
far too many to mention here or describe in any detail. That said, there are a few
with which you should become familiar, as they are likely to become long-term
tools everyone uses. Let’s examine the leading products broken down by category.
MongoDB
CouchDB
Riak
Redis
Voldemort
Cassandra
HBase
Language
C++
Erlang
Erlang
C++
Java
Java
Java
License
AGPL
Apache
Apache
BSD
Apache
Apache
Apache
Model
Document
Document
Key/value
Key/value
Key/value
Wide Column
Wide Column
Protocol
BSON
HTTP/REST HTTP/REST
or TCP/
Protobufs
TCP
TCP/Thrift
HTTP/REST
or TPC/Thrift
Storage
Memory
mapped
b-trees
COW-BTree
Pluggable:
InnoDB,
­LevelDB,
Bitcask
In
memory,
snapshot
to disk
Pluggable:
BDB, MySQL,
in-memory
Memtable/
SSTable
HDFS
Dynamo
Dynamo
Dynamo
BigTable,
Dynamo
BigTable
Inspiration
Search
Yes
No
Yes
No
No
Yes
Yes
MapReduce
Yes
No
Yes
No
No
Yes
Yes
10 ;login: VOL. 3 6, NO. 5
Here are a few company use cases where NoSQL is a critical component of the data
architecture.
Facebook: HBase for Messages
When Facebook decided to expand its messaging services infrastructure [8]
to encompass email, text, instant messages, and more, it knew it had to have a
distributed fault-tolerant database able to manage petabytes of messages and a
write-dominated workload. It needed something to handle their existing service
of “over 350 million users sending over 15 billion person-to-person messages per
month” and a chat service that “supports over 300 million users who send over 120
billion messages per month,” and to allow plenty of room for growth. Although they
considered MySQL, a solution with which they have extensive experience [9], they
created Cassandra [10]—an open source project combining elements of Google’s
BigTable [2] and Amazon’s Dynamo [1] designs—and Apache HBase, a distributed
database that closely mimics Google’s BigTable implementation and is tightly
integrated with Apache Hadoop and ZooKeeper. Facebook favored the strong consistency model, automatic failover, load-balancing, compression, and MapReduce
support of HBase for their new production messaging service.
Craigslist: MongoDB for Archived Postings
Craigslist recently migrated over 2 billion archived postings from its MySQL
clusters into a set of replicated MongoDB servers [11]. They still use MySQL for all
active postings on their site, but these days when you log in and review old posts
that are now expired, you are accessing their new MongoDB-based service. For
Craigslist, scalability and reliability were central requirements, but one of the
more interesting features they gained by going to MongoDB was schema flexibility.
Their MySQL databases were so large that any schema change (ALTER TABLE)
would take around two months to complete across the replicated database set.
Contrast that with MongoDB, where data is stored as JSON documents with no
schema enforcement at all. By separating archived postings from live postings,
Craigslist simplified their architecture and made it easier to change their production schema as requirements changed.
Conclusions and Advice
The SQL vs. NoSQL debate will continue, and both sides will benefit from this
competition in a market that was stagnant for far too long. I think that as the dust
settles it will become evident that NoSQL solutions will work alongside SQL solutions, each doing what they do best. Facebook, Twitter, and many other companies
are integrating NoSQL databases into their infrastructure right alongside SQL
databases. Each has its strengths and weaknesses; neither will entirely displace
the other. Some future SQL databases may start to take on features only found in
NoSQL, such as elasticity and an ability to scale out to large amounts of commodity hardware. The demand for SQL will not go away anytime soon, nor will
the reality of today’s more distributed, virtualized, and commodity-based IT
infrastructure. As more companies begin capturing and analyzing an increasing
amount of data about their customers, the need for databases that can efficiently
manage and analyze that kind of data will only grow.
The key to making a good decision in this market is to remain as objective and
open-minded as possible while evaluating these products and talking to their ven-
;login: O C TO B E R 20 11 NoSQL 11
dors. Recognize that this is a new market, and that right now the market leaders
are working hard to cement their positions. Use that to your advantage by asking
hard questions and expecting detailed answers. Never decide on a solution without
having first put that vendor’s product through rigorous testing using your data on
your systems. Be sure to interact with and learn more about the vendor; they are
all interested in building relationships with you and will generally go out of their
way to help you with their product. Spend some time understanding their particular views on eventual consistency, durability, CAP, and replication, and be certain
you understand how those tradeoffs will impact your design. Make sure that they
support enough tools to meet your needs today and tomorrow. Look for a solid community to bolster your development. In the end, be sure that the solution you pick
is going to support you as your product matures. Make sure the vendor has shared
their road map and explained how they will help you move from one version to the
next, because this area of technology is far from static.
References
[1] Amazon Dynamo: http://www.allthingsdistributed.com/2007/10/amazons
_dynamo.html.
[2] Google BigTable: http://labs.google.com/papers/bigtable.html.
[3] LinkedIn Voldemort: http://project-voldemort.com/.
[4] Twitter FlockDB: http://engineering.twitter.com/2010/05/introducing-flockdb
.html.
[5] Facebook Cassandra: http://cassandra.apache.org/.
[6] Yahoo! PNUTS: http://www.brianfrankcooper.net/pubs/pnuts.pdf.
[7] CAP theorem: http://en.wikipedia.org/wiki/CAP_theorem.
[8] The Underlying Technology of Messages: https://www.facebook.com/notes
/facebook-engineering/the-underlying-technology-of-messages/454991608919.
[9] MySQL at Facebook: http://www.facebook.com/MySQLatFacebook.
[10] Note on creating Cassandra: https://www.facebook.com/note.php?note_id
=24413138919.
[11] Jeremy Zawodny, “Lessons Learned from Migrating 2+ Billion Documents at
Craigslist”: http://www.10gen.com/presentation/mongosf2011/craigslist.
12 ;login: VOL. 3 6, NO. 5
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

advertisement