Gossiper exchanges data with its peers, caching the buckets. When a node leaves the cluster, it now (after BUG-3128) remotes buckets for disconnected nodes.

It does not handle the case of a hung or terminated peer Gossiper, in which case it will leave (and propagate) stale Buckets.

To handle this case Gossiper needs to install a remote death watch, http://doc.akka.io/docs/akka/2.4/scala/remoting.html#Watching_Remote_Actors, to get notified when the advertising actor dies. When such an event occurs, it needs to invalidate the corresponding remote bucket and fire a corresponding message to RpcRegistry.

That way remote RPCs will be correctly unregistered and any requests to those RPCs will fail-fast instead of timing out.