Found in org.jgroups.protocols.pbcast.GMS in a scenario where members were killed and restarted very frequently and one member is persistently running. The memory consumption icreased very quickly until OutOfMemory.

Our analysis with JProfiler leads to the following assumption:
The number of org.jgroups.stack.IpAddress held by merge_ack_collector icreases every time a member was (re)started.
These instances were never removed because no MERGE events occured. This is completely clear since there was only one coordinator (the long running member).
If GMS receives no MERGE event it never cleans up the stored IpAddress instances in the merge_ack_collector.