Voldemort is a fully distributed, key/value store. The distribution model is Dynamo/eventual consistency (there are, however, tunable "knobs": you can choose how much of either consistency, availability and partition tolerance you want to give up in favour of the other two). The storage backends are pluggable, but most commonly used ones are BerkeleyDB and read-only (for data built off-line). Latency is low, throughput is high and there isn't a single point of failure. I am presently working full time on a feature that enables additional nodes (and thus capacity) to be added as a cluster is running (and serving live requests).

HBase is a BigTable style columnar style (with support for certain kinds of range queries: see the Google BigTable paper). The distribution model is strong consistency, the backing store is built on top of Hadoop's HDFS. Hadoop's name mode is the single central point / "special" node (although there's work being done by the Hadoop team on having multiple namenodes, afaik).

Cassandra takes yet another approach, mixing parts of Dynamo's consistency/distribution model with parts of BigTable's data model.

The two are different projects, solving different problems. I am a commiter to Voldemort, but I play with/like HBase and Cassandra as well. One size does not fit all :-)