When I enable replication and eviction, it appear that not all nodes are replicated to all hosts. This problem was discovered when clustering modeshape with eviction, and critical nodes were not being properly replicated.

I've modified the clustered-cache quick-start to (hopefully) demonstrate this problem:

https://github.com/cbeer/infinispan-quickstart/tree/replication-eviction-problem/clustered-cache

Node1 creates 100 cache entries (key0 -> key99). When eviction is disabled, the final cache size on Node0 is 100. When eviction is enabled, the final cache size is 78.

This seems suspiciously similar to ISPN-2712.