Currently when we delete a cluster at runtime, we don't correctly raise member change callbacks, so the hosts will stick around. Need to look at this entire flow and fix/test.