If the client has dropped the connection before the RPC is processed, RPC server doesn't need to process the RPC call. We have encountered issues where bad applications can bring down the NN. https://issues.apache.org/jira/i#browse/Hadoop-9640 tries to address that. When this occurs, NN's RPC queues are filled up with client requests and DN requests, sometimes we want to stop the flooding by stopping the bad applications and/or DNs. Some RPC processing like DatanodeProtocol::blockReport could take couple hundred milliseconds. So it is worthwhile to have NN skip the RPC calls if DNs have been stopped.