Cleaning the application may cost much time at worker, then it will block that the worker send heartbeats master because the worker is extend ThreadSafeRpcEndpoint. If the heartbeat from a worker is blocked by the message ApplicationFinished, master will think the worker is dead. If the worker has a driver, the driver will be scheduled by master again. So I think it is the bug on spark. It may solve this problem by the followed suggests：
1. It had better put the cleaning the application in a single asynchronous thread like 'cleanupThreadExecutor'. Thus it won't block other rpc messages like SendHeartbeat;
2. It had better not receive the heartbeat master by receive method. Because any other rpc message may block the receive method. Then worker won't receive the heartbeat message timely. So it had better send the heartbeat master at an asynchronous timing thread .