MR-7178.md

MR-7178

Scenario

1. Shutdown command comes while YarnChild is initializing

2. line#1 is crash point, so we inject shutdown after it.

   YarnChild.taskid = task.getTaskID();
   try{
    logSyncer = TaskLog.createLogSyncer();//line 168
    ....
    taskFinal.run(job, umbilical); //line 178
   }catch (Exception exception) {//line 187
     LOG.warn("Exception running child : "
      + StringUtils.stringifyException(exception));
      .................
      task.taskCleanup(umbilical);// line 200
   }
   

3. But line168 throw exception:"Shutdown in progress, cannot add a shutdownHook".

4. Line 187 will catch this exception and do clean up(line 200), code line 200 will use task.committer without null check, hence NPE happens

Trigger analysis

1. YarnChild.taskid is the meta-info

2. And it is written, we should crash the YarnChild node, but due to YarnChild.taskid is already on YarnChild node, so crash should change to shutdown.