Data loss caused by a single node which had a long GC pause

There is a data loss scenario that can happen frequently. If a node is dropped from a cluster but the node is actually alive (the java process persists with its heap contents, like by a long GC pause), it rejoins to the cluster soon and override the cluster topology because the dropped node has a topology with the largest size.

For example, suppose a cluster of 3 nodes; A, B and C. C had a long GC pause and is dropped from the cluster. A and B form a new cluster with size 2. When C backs to the cluster, it overrides the topology of size 2 because C remembers a topology when its size was 3. Some of updates when the size was 2 are not accessible any more.

All versions of JDG 6, JDG 7 and Infinispan 9 are affected.