The replication statistics exposed by RpcManagerImpl are too low-level. For instance, in non-transactional caches, the times of the RPCs from the primary to the backups are counted both on the primary and on the originator (as part of the originator-to-primary RPC). They also aren't split by operation, like the CacheMgmtInterceptor statistics, which makes them even less useful.

The triangle algorithm changes things again, but only for distributed caches, so there is now one more way of interpreting the RpcManagerImpl statistics. It would be better to remove them altogether.