We're testing cassandra for integration with indextank. In this first try, we're creating one column family for each user. In practice, on the first run and for the first few documents (a few 100s), a new CF is created, and a document is immediately added to it. A few (up to 50) requests of this type are issued in parallel (for different column families).
The end result, and quite repeatable, is having the cluster split with different schema versions, and they never agree.