It looks like HTableMultiplexer has a couple of issues.
Upon failing to send a Put to the appropriate RS, the Put is re-queued back into the system. Normally this is fine as such an exception is transient and the Put would eventually succeed. However, in the case where the Put was rejected because of a NotServingRegionException (e.g. split, balance, merge), the re-queuing of the Put will end up using the same cached HRegionLocation. This means that the Put will just be repeatedly sent back to the same RS over and over again, eventually being dropped on the floor. Need to invalidate the location cache (or make sure we refresh it) when we re-queue the Put.
The internal ClusterConnection is also leaked. If a user creates many HTableMultiplexers, they'll eventually run into issues (memory, zk connections, etc) because they'll never get cleaned up. HTableMultiplexer needs a close method.