it seems cassandra, mongodb and couch have gotten the lions share of the attention. of the three I would say couch is the closest to riak. Both are native erlang apps, speak http natively and use map/reduce as a primary mechanism of retrieving data.

riak is just about 1 years old from an open source perspective but is in use in a few production environments and had been in development a few yrs before going open source, afaik. What I like most about riak is the 'no special node' feature and the very close adherence to amazon's dynamo lineage. easy scale up - and easy scale down. adherence to dynamo also means you can fine tune your r/w/n values at a granular level. beyond that, the fact that you can do virtually everything there is to do in riak over http is fantastic. other access pathways are erlang client and protobuffs. another major feature of riak is that you can link keys together with link headers. admin wise, riak is very simple to get up and running either in a dev environment or production.

the drawbacks against riak are that there is no ordering of keys, no select * keys, no select range. listing all keys is expensive. you almost have to keep your own internal index of keys if you want to maintain performance over m/r jobs in any sort of realtime fashion.

im glossing over lots of finer points but if you have specific questions the irc is well manned and so is the mailing list.