I recently attended the NoSQL conference in London and think I have a better idea now how to answer the original question. I also wrote a blog post, and there are a couple of other good ones.
Key points:
- We have accumulated probably 30 years knowledge of adminstering relational databases, so shouldn’t replace them without careful consideration; non-relational data stores are less mature than relational ones, and so are inherently more risky to adopt
- There are different types of non-relational data store; some are key-value stores, some are document stores, some are graph databases
- You could use a hybrid approach, e.g. a combination of RDBMS and graph data store for a social software site
- Document data stores (e.g. CouchDB and MongoDB) are probably the closest to relational databases and provide a JSON data structure with all the fields presented hierarchically which avoids having to do table joins, and (some might argue) is an improvement on the traditional object-relational mapping that most applications currently use
- Non-relational databases support replication (including master-master); relational databases support replication too but it may not be as comprehensive as the non-relational option
- Very large sites such as Twitter, Digg and Facebook use Cassandra, which is built from the ground up to support clustering
- Relational databases are probably suitable for 90% of cases
In summary, consensus seems to be “proceed with caution”.