Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
2.0.0-alpha
-
None
-
None
Description
We saw this problem migrating applications to MapReduceV2:
Our applications use hadoop counters extensively (1000+ counters for certain jobs). While this may not be one of recommended best practices in hadoop, the real issue here is reliability of the framework when applications exceed counter limits.
The hadoop servers (yarn, history server) were originally brought up with mapreduce.job.counters.max=1000 under core-site.xml
We then ran map-reduce job under an application using its own job specific overrides, with mapreduce.job.counters.max=10000
All the tasks for the job finished successfully; however the overall job still failed due to AM encountering exceptions as:
2012-07-12 17:31:43,485 INFO [AsyncDispatcher event handler] org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl: Num completed Tasks : 712012-07-12 17:31:43,502 FATAL [AsyncDispatcher event handler] org.apache.hadoop.yarn.event.AsyncDispatcher: Error in dispatcher threa dorg.apache.hadoop.mapreduce.counters.LimitExceededException: Too many counters: 1001 max=1000 at org.apache.hadoop.mapreduce.counters.Limits.checkCounters(Limits.java:58) at org.apache.hadoop.mapreduce.counters.Limits.incrCounters(Limits.java:65) at org.apache.hadoop.mapreduce.counters.AbstractCounterGroup.addCounter(AbstractCounterGroup.java:77) at org.apache.hadoop.mapreduce.counters.AbstractCounterGroup.addCounterImpl(AbstractCounterGroup.java:94) at org.apache.hadoop.mapreduce.counters.AbstractCounterGroup.findCounter(AbstractCounterGroup.java:105) at org.apache.hadoop.mapreduce.counters.AbstractCounterGroup.incrAllCounters(AbstractCounterGroup.java:202) at org.apache.hadoop.mapreduce.counters.AbstractCounters.incrAllCounters(AbstractCounters.java:337) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.constructFinalFullcounters(JobImpl.java:1212) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.mayBeConstructFinalFullCounters(JobImpl.java:1198) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.createJobFinishedEvent(JobImpl.java:1179) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.logJobHistoryFinishedEvent(JobImpl.java:711) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.checkJobCompleteSuccess(JobImpl.java:737) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl$TaskCompletedTransition.checkJobForCompletion(JobImpl.java:1360) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl$TaskCompletedTransition.transition(JobImpl.java:1340) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl$TaskCompletedTransition.transition(JobImpl.java:1323) at org.apache.hadoop.yarn.state.StateMachineFactory$MultipleInternalArc.doTransition(StateMachineFactory.java:380) at org.apache.hadoop.yarn.state.StateMachineFactory.doTransition(StateMachineFactory.java:298) at org.apache.hadoop.yarn.state.StateMachineFactory.access$300(StateMachineFactory.java:43) at org.apache.hadoop.yarn.state.StateMachineFactory$InternalStateMachine.doTransition(StateMachineFactory.java:443) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.handle(JobImpl.java:666) at org.apache.hadoop.mapreduce.v2.app.job.impl.JobImpl.handle(JobImpl.java:113) at org.apache.hadoop.mapreduce.v2.app.MRAppMaster$JobEventDispatcher.handle(MRAppMaster.java:890) at org.apache.hadoop.mapreduce.v2.app.MRAppMaster$JobEventDispatcher.handle(MRAppMaster.java:886) at org.apache.hadoop.yarn.event.AsyncDispatcher.dispatch(AsyncDispatcher.java:125) at org.apache.hadoop.yarn.event.AsyncDispatcher$1.run(AsyncDispatcher.java:74) at java.lang.Thread.run(Thread.java:662) 2012-07-12 17:31:43,502 INFO [AsyncDispatcher event handler] org.apache.hadoop.yarn.event.AsyncDispatcher: Exiting, bbye..2012-07-12 17:31:43,503 INFO [Thread-1] org.apache.had
The overall job failed, and the job history wasn't accessible either at the end of the job (didn't show up in job history server).
We were able to workaround the issue by changing to higher limits in core-site.xml and restarting yarn servers. However that forced us to increase the counters global limit to be as high as possible use by any individual application, which is hard to predict.
The original job then succeeded with new global limits.
However, since we didn't restart the job history server, it was unable to display job history page for the successful job altogether as it still hit counter exceeded exception. Restart of job history server finally got the application available under job history.
I'll also attach AM logs to help debug the issue
Attachments
Attachments
Issue Links
- depends upon
-
MAPREDUCE-4443 MR AM and job history server should be resilient to jobs that exceed counter limits
- Patch Available
- relates to
-
MAPREDUCE-5875 Make Counter limits consistent across JobClient, MRAppMaster, and YarnChild
- Closed