We're shadowing some production traffics to a Rocksandra cluster ( https://github.com/Instagram/cassandra/tree/rocks_3.0), the P99 latency is significantly improved (about 6x for read, 12x for write). Here are the test details:
https://docs.google.com/document/d/1cEM8ZqB5tOYVdsh1LpqSZ-eLasumWfzn_Tk_lDjxOTk/ Thanks, Jay