When starting many instances in parallel, it often happens that the node does not detect its neighborhood very well and this results in many subclusters, merging views etc.

Merging two available partitions has undefined results (AFAIK). While we can expect that there are no requests to the cluster from the application ¹, Infinispan itself uses some caches to store internal information (HotRod routing, Protobuf etc...). It would be better if the available-available merge would provide hooks for rebuilding this info.

¹) Being able to start the cluster with reads/writes disabled and enable them only when the cache has expected number of members would be convenient, too.