Replication

Before you start: For security reasons, CouchDB is by default configured to listen to localhost/127.0.0.1 only. In order to replicate to a remote CouchDB server, the remote CouchDB has to bound to 0.0.0.0, the source server can remain private on 127.0.0.1. Find the option

bind_address
in the http section of the Futon configuration page.

This chapter introduces CouchDB’s world-class replication system. Replication synchronizes two copies of the same database, allowing users to have low latency access to data no matter where they are. These databases can live on the same server or on two different servers—CouchDB doesn’t make a distinction. If you change one copy of the database, replication will send these changes to the other copy.

Replication is a one-off operation: you send an HTTP request to CouchDB that includes a source and a target database, and CouchDB will send the changes from the source to the target. That is all. Granted, calling something world-class and then only needing one sentence to explain it does seem odd. But part of the reason why CouchDB’s replication is so powerful lies in its simplicity.

Let’s see what replication looks like:

POST /_replicate HTTP/1.1
{"source":"database","target":"http://example.org/database"} -H "Content-Type: application/json"

This call sends all the documents in the local database database to the remote database http://example.org/database. A database is considered “local” when it is on the same CouchDB instance you send the POST /_replicate HTTP request to. All other instances of CouchDB are “remote.”

If you want to send changes from the target to the source database, you just make the same HTTP requests, only with source and target database swapped. That is all.

POST /_replicate HTTP/1.1
Content-Type: application/json
{"source":"http://example.org/database","target":"database"}

A remote database is identified by the same URL you use to talk to it. CouchDB replication works over HTTP using the same mechanisms that are available to you. This example shows that replication is a unidirectional process. Documents are copied from one database to another and not automatically vice versa. If you want bidirectional replication, you need to trigger two replications with source and target swapped.

The Magic

When you ask CouchDB to replicate one database to another, it will go and compare the two databases to find out which documents on the source differ from the target and then submit a batch of the changed documents to the target until all changes are transferred. Changes include new documents, changed documents, and deleted documents. Documents that already exist on the target in the same revision are not transferred; only newer revisions are.

Databases in CouchDB have a sequence number that gets incremented every time the database is changed. CouchDB remembers what changes came with which sequence number. That way, CouchDB can answer questions like, “What changed in database A between sequence number 212 and now?” by returning a list of new and changed documents. Finding the differences between databases this way is an efficient operation. It also adds to the robustness of replication.

CouchDB views use the same mechanism when determining when a view needs updating and which documents to replicate. You can use this to build your own solutions as well.

You can use replication on a single CouchDB instance to create snapshots of your databases to be able to test code changes without risking data loss or to be able to refer back to older states of your database. But replication gets really fun if you use two or more different computers, potentially geographically spread out.

With different servers, potentially hundreds or thousands of miles apart, problems are bound to happen. Servers crash, network connections break off, things go wrong. When a replication process is interrupted, it leaves two replicating CouchDBs in an inconsistent state. Then, when the problems are gone and you trigger replication again, it continues where it left off.

Simple Replication with the Admin Interface

You can run replication from your web browser using Futon, CouchDB’s built-in administration interface. Start CouchDB and open your browser to http://127.0.0.1:5984/_utils/. On the righthand side, you will see a list of things to visit in Futon. Click on “Replication.”

Futon will show you an interface to start replication. You can specify a source and a target by either picking a database from the list of local databases or filling in the URL of a remote database.

Click on the Replicate button, wait a bit, and have a look at the lower half of the screen where CouchDB gives you some statistics about the replication run or, if an error occurred, an explanatory message.

Congratulations—you ran your first replication.

Replication in Detail

So far, we’ve skipped over the result from a replication request. Now is a good time to look at it in detail. Here’s a nicely formatted example:

{
  "ok": true,
  "source_last_seq": 10,
  "session_id": "c7a2bbbf9e4af774de3049eb86eaa447",
  "history": [
    {
      "session_id": "c7a2bbbf9e4af774de3049eb86eaa447",
      "start_time": "Mon, 24 Aug 2009 09:36:46 GMT",
      "end_time": "Mon, 24 Aug 2009 09:36:47 GMT",
      "start_last_seq": 0,
      "end_last_seq": 1,
      "recorded_seq": 1,
      "missing_checked": 0,
      "missing_found": 1,
      "docs_read": 1,
      "docs_written": 1,
      "doc_write_failures": 0,
    }
  ]
}

The "ok": true part, similar to other responses, tells us everything went well. source_last_seq includes the source’s update_seq value that was considered by this replication. Each replication request is assigned a session_id, which is just a UUID; you can also talk about a replication session identified by this ID.

The next bit is the replication history. CouchDB maintains a list of history sessions for future reference. The history array is currently capped at 50 entries. Each unique replication trigger object (the JSON string that includes the source and target databases as well as potential options) gets its own history. Let’s see what a history entry is all about.

The session_id is recorded here again for convenience. The start and end time for the replication session are recorded. The _last_seq denotes the update_seqs that were valid at the beginning and the end of the session. recorded_seq is the update_seq of the target again. It’s different from end_last_seq if a replication process dies in the middle and is restarted. missing_checked is the number of docs on the target that are already there and don’t need to be replicated. missing_found is the number of missing documents on the source.

The last three—docs_read, docs_written, and doc_write_failures—show how many documents we read from the source, wrote to the target, and how many failed. If all is well, _read and _written are identical and doc_write_failures is 0. If not, you know something went wrong during replication. Possible failures are a server crash on either side, a lost network connection, or a validate_doc_update function rejecting a document write.

One common scenario is triggering replication on nodes that have admin accounts enabled. Creating design documents is restricted to admins, and if the replication is triggered without admin credentials, writing the design documents during replication will fail and be recorded as doc_write_failures. If you have admins, be sure to include the credentials in the replication request:

> curl -X POST http://127.0.0.1:5984/_replicate  -d '{"source":"http://example.org/database", "target":"http://admin:password@127.0.0.1:5984/database"}' -H "Content-Type: application/json"

Continuous Replication

Now that you know how replication works under the hood, we share a neat little trick. When you add "continuous":true to the replication trigger object, CouchDB will not stop after replicating all missing documents from the source to the target. It will listen on CouchDB’s _changes API (see Chapter 20, Change Notifications) and automatically replicate over any new docs as they come into the source to the target. In fact, they are not replicated right away; there’s a complex algorithm determining the ideal moment to replicate for maximum performance. The algorithm is complex and is fine-tuned every once in a while, and documenting it here wouldn’t make much sense.

> curl -X POST http://127.0.0.1:5984/_replicate -d '{"source":"db", "target":"db-replica", "continuous":true}' -H "Content-Type: application/json"

The replicator database

Requests posted to /_replicate/ trigger a single replication operation, or if the continuous flag is set, a single replication thread that will continue replicating until the server dies.

Since CouchDB 1.1.0, permanent continuous replications that survive a server restart without you having to do anything can be defined by inserting documents in the replicator database (which by default is named _replicator). These documents have the same syntax as the JSON objects posted to /_replicate/.

That’s It?

Replication is the foundation on which the following chapters build. Make sure you have understood this chapter. If you don’t feel comfortable yet, just read it again and play around with the replication interface in Futon.

We haven’t yet told you everything about replication. The next chapters show you how to manage replication conflicts (see Chapter 17, Conflict Management), how to use a set of synchronized CouchDB instances for load balancing (see Chapter 18, Load Balancing), and how to build a cluster of CouchDBs that can handle more data or write requests than a single node (see Chapter 19, Clustering).