Eventual Consistency

In the previous chapter, we saw that CouchDB’s flexibility allows us to evolve our data as our applications grow and change. In this chapter, we’ll explore how working “with the grain” of CouchDB promotes simplicity in our applications and helps us naturally build scalable, distributed systems.

Working with the Grain

A distributed system is a system that operates robustly over a wide network. A particular feature of network computing is that network links can potentially disappear, and there are plenty of strategies for managing this type of network segmentation. CouchDB differs from others by accepting eventual consistency, as opposed to putting absolute consistency ahead of raw availability like RDBMS or Paxos. What these systems have in common is an awareness that data acts differently when many people are accessing it simultaneously. Their approaches differ when it comes to which aspects of consistency, availability, or partition tolerance they prioritize.

Engineering distributed systems is tricky. Many of the caveats and “gotchas” you will face over time aren’t immediately obvious. We don’t have all the solutions, and CouchDB isn’t a panacea, but when you work with CouchDB’s grain rather than against it, the path of least resistance leads you to naturally scalable applications.

Of course, building a distributed system is only the beginning. A website with a database that is available only half the time is next to worthless. Unfortunately, the traditional relational database approach to consistency makes it very easy for application programmers to rely on global state, global clocks, and other high availability no-nos, without even realizing that they’re doing so. Before examining how CouchDB promotes scalability, we’ll look at the constraints faced by a distributed system. After we’ve seen the problems that arise when parts of your application can’t rely on being in constant contact with each other, we’ll see that CouchDB provides an intuitive and useful way for modeling applications around high availability.

The CAP Theorem

The CAP theorem describes a few different strategies for distributing application logic across networks. CouchDB’s solution uses replication to propagate application changes across participating nodes. This is a fundamentally different approach from consensus algorithms and relational databases, which operate at different intersections of consistency, availability, and partition tolerance.

The CAP theorem, shown in Figure 1, “The CAP theorem”, identifies three distinct concerns:

Consistency
All database clients see the same data, even with concurrent updates.
Availability
All database clients are able to access some version of the data.
Partition tolerance
The database can be split over multiple servers.

Pick two.

Figure 1. The CAP theorem

When a system grows large enough that a single database node is unable to handle the load placed on it, a sensible solution is to add more servers. When we add nodes, we have to start thinking about how to partition data between them. Do we have a few databases that share exactly the same data? Do we put different sets of data on different database servers? Do we let only certain database servers write data and let others handle the reads?

Regardless of which approach we take, the one problem we’ll keep bumping into is that of keeping all these database servers in sync. If you write some information to one node, how are you going to make sure that a read request to another database server reflects this newest information? These events might be milliseconds apart. Even with a modest collection of database servers, this problem can become extremely complex.

When it’s absolutely critical that all clients see a consistent view of the database, the users of one node will have to wait for any other nodes to come into agreement before being able to read or write to the database. In this instance, we see that availability takes a backseat to consistency. However, there are situations where availability trumps consistency:

Each node in a system should be able to make decisions purely based on local state. If you need to do something under high load with failures occurring and you need to reach agreement, you’re lost. If you’re concerned about scalability, any algorithm that forces you to run agreement will eventually become your bottleneck. Take that as a given.

—Werner Vogels, Amazon CTO and Vice President

If availability is a priority, we can let clients write data to one node of the database without waiting for other nodes to come into agreement. If the database knows how to take care of reconciling these operations between nodes, we achieve a sort of “eventual consistency” in exchange for high availability. This is a surprisingly applicable trade-off for many applications.

Unlike traditional relational databases, where each action performed is necessarily subject to database-wide consistency checks, CouchDB makes it really simple to build applications that sacrifice immediate consistency for the huge performance improvements that come with simple distribution.

Local Consistency

Before we attempt to understand how CouchDB operates in a cluster, it’s important that we understand the inner workings of a single CouchDB node. The CouchDB API is designed to provide a convenient but thin wrapper around the database core. By taking a closer look at the structure of the database core, we’ll have a better understanding of the API that surrounds it.

The Key to Your Data

At the heart of CouchDB is a powerful B-tree storage engine. A B-tree is a sorted data structure that allows for searches, insertions, and deletions in logarithmic time. As Figure 2, “Anatomy of a view request” illustrates, CouchDB uses this B-tree storage engine for all internal data, documents, and views. If we understand one, we will understand them all.

Figure 2. Anatomy of a view request

CouchDB uses MapReduce to compute the results of a view. MapReduce makes use of two functions, “map” and “reduce,” which are applied to each document in isolation. Being able to isolate these operations means that view computation lends itself to parallel and incremental computation. More important, because these functions produce key/value pairs, CouchDB is able to insert them into the B-tree storage engine, sorted by key. Lookups by key, or key range, are extremely efficient operations with a B-tree, described in big O notation as O(log N) and O(log N + K), respectively.

In CouchDB, we access documents and view results by key or key range. This is a direct mapping to the underlying operations performed on CouchDB’s B-tree storage engine. Along with document inserts and updates, this direct mapping is the reason we describe CouchDB’s API as being a thin wrapper around the database core.

Being able to access results by key alone is a very important restriction because it allows us to make huge performance gains. As well as the massive speed improvements, we can partition our data over multiple nodes, without affecting our ability to query each node in isolation. BigTable, Hadoop, SimpleDB, and memcached restrict object lookups by key for exactly these reasons.

No Locking

A table in a relational database is a single data structure. If you want to modify a table—say, update a row—the database system must ensure that nobody else is trying to update that row and that nobody can read from that row while it is being updated. The common way to handle this uses what’s known as a lock. If multiple clients want to access a table, the first client gets the lock, making everybody else wait. When the first client’s request is processed, the next client is given access while everybody else waits, and so on. This serial execution of requests, even when they arrived in parallel, wastes a significant amount of your server’s processing power. Under high load, a relational database can spend more time figuring out who is allowed to do what, and in which order, than it does doing any actual work.

Note that modern relational databases avoid locks by implementing MVCC under the hood, but hide it from the end user, requiring them to coordinate concurrent changes of single rows or fields.

Instead of locks, CouchDB uses Multi-Version Concurrency Control (MVCC) to manage concurrent access to the database. Figure 3, “MVCC means no locking” illustrates the differences between MVCC and traditional locking mechanisms. MVCC means that CouchDB can run at full speed, all the time, even under high load. Requests are run in parallel, making excellent use of every last drop of processing power your server has to offer.

Figure 3. MVCC means no locking

Documents in CouchDB are versioned, much like they would be in a regular version control system such as Subversion or Git. If you want to change a value in a document, you create an entire new version of that document and save it over the old one. After doing this, you end up with two versions of the same document, one old and one new.

How does this offer an improvement over locks? Consider a set of requests wanting to access a document. The first request reads the document. While this is being processed, a second request changes the document. Since the second request includes a completely new version of the document, CouchDB can simply append it to the database without having to wait for the read request to finish.

When a third request wants to read the same document, CouchDB will point it to the new version that has just been written. During this whole process, the first request could still be reading the original version.

A read request will always see the most recent snapshot of your database at the time of the beginning of the request.

Validation

As application developers, we have to think about what sort of input we should accept and what we should reject. The expressive power to do this type of validation over complex data within a traditional relational database leaves a lot to be desired. Fortunately, CouchDB provides a powerful way to perform per-document validation from within the database.

CouchDB can validate documents using JavaScript functions similar to those used for MapReduce. Each time you try to modify a document, CouchDB will pass the validation function a copy of the existing document, a copy of the new document, and a collection of additional information, such as user authentication details. The validation function now has the opportunity to approve or deny the update.

By working with the grain and letting CouchDB do this for us, we save ourselves a tremendous amount of CPU cycles that would otherwise have been spent serializing object graphs from SQL, converting them into domain objects, and using those objects to do application-level validation.

Distributed Consistency

Maintaining consistency within a single database node is relatively easy for most databases. The real problems start to surface when you try to maintain consistency between multiple database servers. If a client makes a write operation on server A, how do we make sure that this is consistent with server B, or C, or D? For relational databases, this is a very complex problem with entire books devoted to its solution. You could use multi-primary, primary/replica, partitioning, sharding, write-through caches, and all sorts of other complex techniques.

Incremental Replication

CouchDB's operations take place within the context of a single document. As CouchDB achieves eventual consistency between multiple databases by using incremental replication you no longer have to worry about your database servers being able to stay in constant communication. Incremental replication is a process where document changes are periodically copied between servers. We are able to build what’s known as a shared nothing cluster of databases where each node is independent and self-sufficient, leaving no single point of contention across the system.

Need to scale out your CouchDB database cluster? Just throw in another server.

As illustrated in Figure 4, “Incremental replication between CouchDB nodes”, with CouchDB’s incremental replication, you can synchronize your data between any two databases however you like and whenever you like. After replication, each database is able to work independently.

You could use this feature to synchronize database servers within a cluster or between data centers using a job scheduler such as cron, or you could use it to synchronize data with your laptop for offline work as you travel. Each database can be used in the usual fashion, and changes between databases can be synchronized later in both directions.

Figure 4. Incremental replication between CouchDB nodes

What happens when you change the same document in two different databases and want to synchronize these with each other? CouchDB’s replication system comes with automatic conflict detection and resolution. When CouchDB detects that a document has been changed in both databases, it flags this document as being in conflict, much like they would be in a regular version control system.

This isn’t as troublesome as it might first sound. When two versions of a document conflict during replication, the winning version is saved as the most recent version in the document’s history. Instead of throwing the losing version away, as you might expect, CouchDB saves this as a previous version in the document’s history, so that you can access it if you need to. This happens automatically and consistently, so both databases will make exactly the same choice.

It is up to you to handle conflicts in a way that makes sense for your application. You can leave the chosen document versions in place, revert to the older version, or try to merge the two versions and save the result.

Case Study

Greg Borenstein, a friend and coworker, built a small library for converting Songbird playlists to JSON objects and decided to store these in CouchDB as part of a backup application. The completed software uses CouchDB’s MVCC and document revisions to ensure that Songbird playlists are backed up robustly between nodes.

Songbird is a free software media player with an integrated web browser, based on the Mozilla XULRunner platform. Songbird is available for Microsoft Windows, Apple Mac OS X, Solaris, and Linux.

Let’s examine the workflow of the Songbird backup application, first as a user backing up from a single computer, and then using Songbird to synchronize playlists between multiple computers. We’ll see how document revisions turn what could have been a hairy problem into something that just works.

The first time we use this backup application, we feed our playlists to the application and initiate a backup. Each playlist is converted to a JSON object and handed to a CouchDB database. As illustrated in Figure 5, “Backing up to a single database”, CouchDB hands back the document ID and revision of each playlist as it’s saved to the database.

Figure 5. Backing up to a single database

After a few days, we find that our playlists have been updated and we want to back up our changes. After we have fed our playlists to the backup application, it fetches the latest versions from CouchDB, along with the corresponding document revisions. When the application hands back the new playlist document, CouchDB requires that the document revision is included in the request.

CouchDB then makes sure that the document revision handed to it in the request matches the current revision held in the database. Because CouchDB updates the revision with every modification, if these two are out of sync it suggests that someone else has made changes to the document between the time we requested it from the database and the time we sent our updates. Making changes to a document after someone else has modified it without first inspecting those changes is usually a bad idea.

Forcing clients to hand back the correct document revision is the heart of CouchDB’s optimistic concurrency.

We have a laptop we want to keep synchronized with our desktop computer. With all our playlists on our desktop, the first step is to “restore from backup” onto our laptop. This is the first time we’ve done this, so afterward our laptop should hold an exact replica of our desktop playlist collection.

After editing our Argentine Tango playlist on our laptop to add a few new songs we’ve purchased, we want to save our changes. The backup application replaces the playlist document in our laptop CouchDB database and a new document revision is generated. A few days later, we remember our new songs and want to copy the playlist across to our desktop computer. As illustrated in Figure 6, “Synchronizing between two databases”, the backup application copies the new document and the new revision to the desktop CouchDB database. Both CouchDB databases now have the same document revision.

Figure 6. Synchronizing between two databases

Because CouchDB tracks document revisions, it ensures that updates like these will work only if they are based on current information. If we had made modifications to the playlist backups between synchronization, things wouldn’t go as smoothly.

We back up some changes on our laptop and forget to synchronize. A few days later, we’re editing playlists on our desktop computer, make a backup, and want to synchronize this to our laptop. As illustrated in Figure 7, “Synchronization conflicts between two databases”, when our backup application tries to replicate between the two databases, CouchDB sees that the changes being sent from our desktop computer are modifications of out-of-date documents and helpfully informs us that there has been a conflict.

Recovering from this error is easy to accomplish from an application perspective. Just download CouchDB’s version of the playlist and provide an opportunity to merge the changes or save local modifications into a new playlist.

Figure 7. Synchronization conflicts between two databases

Wrapping Up

CouchDB’s design borrows heavily from web architecture and the lessons learned deploying massively distributed systems on that architecture. By understanding why this architecture works the way it does, and by learning to spot which parts of your application can be easily distributed and which parts cannot, you’ll enhance your ability to design distributed and scalable applications, with CouchDB or without it.

We’ve covered the main issues surrounding CouchDB’s consistency model and hinted at some of the benefits to be had when you work with CouchDB and not against it. But enough theory—let’s get up and running and see what all the fuss is about!