The cassandra driver pools connections automatically, but unlike many JDBC drivers it doesn't cache PreparedStatement handles on the client side. Thus each prepare call is a network round trip, potentially to multiple db nodes. Apps with a fixed set of statements typically declare them in code as fields, but where they are dynamically generated at runtime that's not feasible. Hence we prefer a caching approach. The support bound on cache size will have to be a reasonable guess at this stage. Support for cache size configuration may be worthwhile if customer use shows it's a memory hog.