As seen here the DistributedPubSubMediator removes nodes (and hence potential recipients for messages) from it's registry, when the MemberRemoved event occurs. This might be too late, since it is common behavior, that a gracefully leaving node terminates after it sees it's own MemberRemoved event. But this might happen before this information has converged across the whole cluster, so other nodes might send messages to a gracefully leaving node.

See the following timeline:

If DistributedPubSubMediator would remove nodes from it's registry on exiting then it would be guaranteed, that gracefully leaving nodes are removed from all PubSub in the cluster before getting their own removed event and hence terminating. This way no messages would be lost because of termination/pubsub unregister race.

/cc @ktoso