I have a client comparing speeds with the driver they were using (cql-1.4.0) and the latest DataStax python driver and they're seeing major slowdowns of 2x in their application code and 5x in their simple synthetic tests. However, their use case is fairly unique and they're connecting locally on a single node "cluster".
I went ahead and wrote a short benchmark (code attached) using comparable code to the client and I am seeing similar results:
DataStax with libev:
cql driver over thrift:
ubuntu 14.04 LTS
4 gigs of ram
SSD backed storage on a VM
I've tried this with Apache Cassandra 2.0.6 and DSE 4.5.1 (equivalent Cassandra version 184.108.40.206).