Two points that are specified but somewhat hidden in the article: 1) the 400% number is between Cassandra 0.6 and Cassandra 1.0, 2) the test is CPU bound because the dataset fits very easily in memory.
A lot of awesome work has gone into Cassandra 1.0: many thanks to the Datastax team.
Cassandra is way underrated. It is a great advance in DBRMS, just not relational. I've put it to great use and it's an amazing database system. Searching is badly needed, regular expression searching by column values would make it kick Oracle`s ass anytime.
For an outsider, its hard to understand whats up w/ Cassandra versus HBase. Then you read companies X, Y and Z are moving to HBase, and it almost feels like the project is "just not any good anymore".
EDIT: I think the fallacy of my reasoning is thinking of Apache as a company which wouldn't spend resources on two similar things.
From what I understand HBase is a little more feature complete. It has row locking built in for instance. Cassandra really can't do that by the very nature of how it is designed(fully peer to peer, no zookeeper).
On the other hand Cassandra is really easy to get going, just download a build and double click. Hbase requires zookeeper and hadoop\hdfs before you can get it running as such it is more complicated and the config is a bit more squirrelly. It really is MySQL vs Postgresql all over again.
A lot of awesome work has gone into Cassandra 1.0: many thanks to the Datastax team.