Best practice is to leave the G1 garbage-collector to its default settings (unless a good reason can be found to do otherwise). In the RESTBase Cassandra cluster we ended up with several settings that were operationalized without much in the way of justification. The lone exception, our use of 8MB region sizes, which were proven to avoid allocation of so-called Humongous Objects, which the collector handles in an exceptional (and sub-optimal) manner.

Setting the collector back to its defaults (sans the region size) has been tested in the dev environment with positive results. We should do the same in production.