I'm evaluating a storage platform for an upcoming project and keep coming back to Cassandra. For this project loosing any amount of data is unacceptable. So far we've used a relational database (Microsoft SQL Server), but the data is so varied and large that it has become an issue to store and query.
Is Cassandra robust enough to use as a primary data store? Or should it only be used to mirror existing data to speed up access?
Anecdotally: yes, Twitter, Digg, Ooyala, SimpleGeo, Mahalo, and others are using or moving to Cassandra for a primary data store (http://n2.nabble.com/Cassandra-users-survey-td4040068.html).
Technically: yes; besides supporting replication (including to multiple datacenters), each Cassandra node has an fsync'd commit log to make sure writes are durable; from there writes are turned into SSTables which are immutable until compaction (which combines multiple SSTables to GC old versions). Snapshotting is supported at any time, including automatic snapshot-before-compaction.