[MAPREDUCE-4774] JobImpl does not handle asynchronous task events in FAILED state - ASF JIRA

Voters

Watch issue

Watchers

Create sub-task

Link

Clone

Update Comment Author

Replace String in Comment

Update Comment Visibility

Delete Comments

XML

Word

Printable

JSON

Details

Type: Bug
Status: Closed
Priority: Major
Resolution: Fixed
Affects Version/s: 0.23.3, 2.0.1-alpha
Fix Version/s: 2.0.3-alpha, 0.23.5
Component/s: applicationmaster, mrv2
Labels:
None

Target Version/s:

2.0.3-alpha, 0.23.5

Description

The test org.apache.hadoop.mapred.TestClusterMRNotification.testMR frequently fails in mapred build (e.g. see https://builds.apache.org/job/PreCommit-MAPREDUCE-Build/2988/testReport/junit/org.apache.hadoop.mapred/TestClusterMRNotification/testMR/ , or
https://builds.apache.org/job/PreCommit-MAPREDUCE-Build/2982//testReport/org.apache.hadoop.mapred/TestClusterMRNotification/testMR/).

The test aims to check Job status notifications received through HTTP Servlet. It runs 3 jobs: successfull, killed, and failed.
The test expects the servlet to receive some expected notifications in some expected order. It also tries to test the retry-on-failure notification functionality, so on each 1st notification the servlet answers "400 forcing error", and on each 2nd notification attempt it answers "ok".
In general, the test fails because the actual number and/or type of the notifications differs from the expected.

Investigation shows that actual root cause of the problem is an incorrect job state transition: the 3rd job mapred task fails (by intentionally thrown RuntimeException, see UtilsForTests#runJobFail()), and the state of the task changes from RUNNING to FAILED.
At this point JobEventType.JOB_TASK_ATTEMPT_COMPLETED event is submitted (in method org.apache.hadoop.mapreduce.v2.app.job.impl.TaskImpl.handleTaskAttemptCompletion(TaskAttemptId, TaskAttemptCompletionEventStatus)), and this event gets processed in AsyncDispatcher, but this transition is impossible according to the event transition map (see org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl#stateMachineFactory). This causes the following exception to be thrown upon the event processing:
2012-11-06 12:22:02,335 ERROR [AsyncDispatcher event handler] org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl: Can't handle this event at current state
org.apache.hadoop.yarn.state.InvalidStateTransitonException: Invalid event: JOB_TASK_ATTEMPT_COMPLETED at FAILED
at org.apache.hadoop.yarn.state.StateMachineFactory.doTransition(StateMachineFactory.java:309)
at org.apache.hadoop.yarn.state.StateMachineFactory.access$3(StateMachineFactory.java:290)
at org.apache.hadoop.yarn.state.StateMachineFactory$InternalStateMachine.doTransition(StateMachineFactory.java:454)
at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.handle(JobImpl.java:716)
at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.handle(JobImpl.java:1)
at org.apache.hadoop.mapreduce.v2.app.MRAppMaster$JobEventDispatcher.handle(MRAppMaster.java:917)
at org.apache.hadoop.mapreduce.v2.app.MRAppMaster$JobEventDispatcher.handle(MRAppMaster.java:1)
at org.apache.hadoop.yarn.event.AsyncDispatcher.dispatch(AsyncDispatcher.java:130)
at org.apache.hadoop.yarn.event.AsyncDispatcher$1.run(AsyncDispatcher.java:79)
at java.lang.Thread.run(Thread.java:662)

So, the job gets into state "INTERNAL_ERROR", the job end notification like this is sent:
http://localhost:48656/notification/mapred?jobId=job_1352199715842_0002&jobStatus=ERROR
(here we can see "ERROR" status instead of "FAILED")
After that the notification servlet receives either only "ERROR" notification, or one more notification "ERROR" after "FAILED", which finally causes the test to fail. (Some variation in the test behavior caused by racing conditions because there are many asynchronous processings there, and the test is flaky, in fact).

In any way, it looks like the root cause of the problem is the possibility of the forbidden transition "Invalid event: JOB_TASK_ATTEMPT_COMPLETED at FAILED".
Need an expert advice on how that should be fixed.