Interesting, we're doing similar things with Spark, Cassandra and Mesos. By scaling out Mesos with spot instances for its agents.