By reducing latency for a given system (and that is true not only for storage systems), you affect concurrency and throughput capabilities.
When you want to scale out, there are three distinct issues to deal with: scaling read requests, write requests, and data. Orthogonal to all three and to the items shown in Figures 2 and 3 are many more attributes like reliability or simplicity. You can draw many of these graphs that show how different features or attributes pull into different directions and thus shape the system they describe.
CouchDB is very flexible and gives you enough building blocks to create a system shaped to suit your exact problem. That’s not saying that CouchDB can be bent to solve any problem—CouchDB is no silver bullet—but in the area of data storage, it can get you a long way.
CouchDB replication is one of these building blocks. Its fundamental function is to synchronize two or more CouchDB databases. This may sound simple, but the simplicity is key to allowing replication to solve a number of problems: reliably synchronize databases between multiple machines for redundant data storage; distribute data to a cluster of CouchDB instances that share a subset of the total number of requests that hit the cluster (load balancing); and distribute data between physically distant locations, such as one office in New York and another in Tokyo.
CouchDB replication uses the same REST API all clients use. HTTP is ubiquitous and well understood. Replication works incrementally; that is, if during replication anything goes wrong, like dropping your network connection, it will pick up where it left off the next time it runs. It also only transfers data that is needed to synchronize databases.
A core assumption CouchDB makes is that things can go wrong, like network connection troubles, and it is designed for graceful error recovery instead of assuming all will be well. The replication system’s incremental design shows that best. The ideas behind “things that can go wrong” are embodied in the Fallacies of Distributed Computing:
- The network is reliable.
- Latency is zero.
- Bandwidth is infinite.
- The network is secure.
- Topology doesn’t change.
- There is one administrator.
- Transport cost is zero.
- The network is homogeneous.
Existing tools often try to hide the fact that there is a network and that any or all of the previous conditions don’t exist for a particular system. This usually results in fatal error scenarios when something finally goes wrong. In contrast, CouchDB doesn’t try to hide the network; it just handles errors gracefully and lets you know when actions on your end are required.
Local Data Is King
CouchDB takes quite a few lessons learned from the Web, but there is one thing that could be improved about the Web: latency. Whenever you have to wait for an application to respond or a website to render, you almost always wait for a network connection that isn’t as fast as you want it at that point. Waiting a few seconds instead of milliseconds greatly affects user experience and thus user satisfaction.
What do you do when you are offline? This happens all the time—your DSL or cable provider has issues, or your iPhone, Android, or Blackberry has no bars, and no connectivity means no way to get to your data.
CouchDB can solve this scenario as well, and this is where scaling is important again. This time it is scaling down. Imagine CouchDB installed on phones and other mobile devices that can synchronize data with centrally hosted CouchDBs when they are on a network. The synchronization is not bound by user interface constraints like subsecond response times. It is easier to tune for high bandwidth and higher latency than for low bandwidth and very low latency. Mobile applications can then use the local CouchDB to fetch data, and since no remote networking is required for that, latency is low by default.
Can you really use CouchDB on a phone? Erlang, CouchDB’s implementation language has been designed to run on embedded devices magnitudes smaller and less powerful than today’s phones.
At the end, this is what the CouchDB defined.
In my point of view:
CouchDB (from their website)
- A document database server, accessible via a RESTful JSON API. Generally, relational databases aren’t simply accessed via REST services, but require a much more complex SQL API. Often these API’s (JDBC, ODBC, etc.) are quite complex. REST is quite simple.
- Ad-hoc and schema-free with a flat address space. Relational databases have complex, fixed schema. You define tables, columns, indexes, sequences, views and other stuff. Couch doesn’t require this level of complex, expensive, fragile advanced planning.
- Distributed, featuring robust, incremental replication with bi-directional conflict detection and management. Some SQL commercial products offer this. Because of the SQL API and the fixed schemas, this is complex, difficult and expensive. For Couch, it appears simple and inexpensive.
So. Why CouchDB?
- REST is simpler than JDBC or ODBC.
- No Schema is simpler than Schema.
- Distributed in a way that appears simple and inexpensive.