Yuqi Wang commented on YARN-6959:

The whole pipeline was that:
Step0. AM sent heartbeats to RM.
Step1. AM process crashed with exitcode 15 without unregister to RM.
Step2-a. NM told RM the AM container has completed.
Step2-b. The heartbeats sent in step0, was processing by RM between MARK1 and 
Step3. RM switched to the new attempt.
Step4. The heartbeats record request from previous AM into current attempt.

So, it is possible.

> RM may allocate wrong AM Container for new attempt
> --------------------------------------------------
>                 Key: YARN-6959
>                 URL: https://issues.apache.org/jira/browse/YARN-6959
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: capacity scheduler, fairscheduler, scheduler
>    Affects Versions: 2.7.1
>            Reporter: Yuqi Wang
>            Assignee: Yuqi Wang
>              Labels: patch
>             Fix For: 2.7.1, 3.0.0-alpha4
>         Attachments: YARN-6959.001.patch, YARN-6959.002.patch, 
> YARN-6959.003.patch, YARN-6959.004.patch, YARN-6959.005.patch, 
> YARN-6959-branch-2.7.001.patch, YARN-6959.yarn_nm.log.zip, 
> YARN-6959.yarn_rm.log.zip
> *Issue Summary:*
> Previous attempt ResourceRequest may be recorded into current attempt 
> ResourceRequests. These mis-recorded ResourceRequests may confuse AM 
> Container Request and Allocation for current attempt.
> *Issue Pipeline:*
> {code:java}
> // Executing precondition check for the incoming attempt id.
> ApplicationMasterService.allocate() ->
> scheduler.allocate(attemptId, ask, ...) ->
> // Previous precondition check for the attempt id may be outdated here, 
> // i.e. the currentAttempt may not be the corresponding attempt of the 
> attemptId.
> // Such as the attempt id is corresponding to the previous attempt.
> currentAttempt = scheduler.getApplicationAttempt(attemptId) ->
> // Previous attempt ResourceRequest may be recorded into current attempt 
> ResourceRequests
> currentAttempt.updateResourceRequests(ask) ->
> // RM may allocate wrong AM Container for the current attempt, because its 
> ResourceRequests
> // may come from previous attempt which can be any ResourceRequests previous 
> AM asked
> // and there is not matching logic for the original AM Container 
> ResourceRequest and 
> // the returned amContainerAllocation below.
> AMContainerAllocatedTransition.transition(...) ->
> amContainerAllocation = scheduler.allocate(currentAttemptId, ...)
> {code}
> *Patch Correctness:*
> Because after this Patch, RM will definitely record ResourceRequests from 
> different attempt into different objects of 
> SchedulerApplicationAttempt.AppSchedulingInfo.
> So, even if RM still record ResourceRequests from old attempt at any time, 
> these ResourceRequests will be recorded in old AppSchedulingInfo object which 
> will not impact current attempt's resource requests and allocation.
> *Concerns:*
> The getApplicationAttempt function in AbstractYarnScheduler is so confusing, 
> we should better rename it to getCurrentApplicationAttempt. And reconsider 
> whether there are any other bugs related to getApplicationAttempt.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org

Reply via email to