Looks like 0.14 is a great release, from the notes. I'm setting up a Riak cluster tomorrow to evaluate performance as an alternative to our current HBase-based setup, so the timing on the release is perfect for me.
FWIW, if you're looking at HBase alternatives because of stability or reliability concerns, Cassandra is probably going to be an easy move because of the common BigTable datamodel ancestor. (With some notable improvements, e.g., http://www.riptano.com/blog/whats-new-cassandra-07-secondary...)
Stability and reliability have dramatically improved in HBase over time (2 years ago, those were definitely concerns I had). When we began using HBase in mid-2007, it seemed like the best (read: only viable) option for our use case. Now, there are comparably lots of options. Without much testing so far, Riak's fully distributed (non-master/slave) architecture seems appealing..
For any SQL/NoSQL system, the more RAM, the better. They need as much RAM as possible to cache the data.
If you observed cassandra consumes a lot of RAM even with small dataset, you should tune the GC settings. Besides it, JVM itself consumes a lot of RAM.
If you're running a "real" cluster (i.e., where you actually need multiple machines instead of just one big one running postgresql), then yes, the recommendations there are realistic -- not because cassandra is inefficient but because you will ipso facto be dealing with a large dataset.
Cassandra does scale down (I know at least one site that went live running Cassandra on a 256MB VM) but that's mostly of academic interest to the audience looking for a scalable database. Hence the recommendations on that page.
