Jian He commented on YARN-2853:

Instead, we could get rid of the killing state completely and let app stay at 
the original state and change RMApp to handle attempt_killed state at each 
possible state.  this way, we could avoid race condition like this.  I'll file 
a separate  jira to do this.

> Killing app may hang while AM is unregistering
> ----------------------------------------------
>                 Key: YARN-2853
>                 URL: https://issues.apache.org/jira/browse/YARN-2853
>             Project: Hadoop YARN
>          Issue Type: Bug
>            Reporter: Jian He
>            Assignee: Jian He
>         Attachments: YARN-2853.1.patch
> When killing an app, app first moves to KILLING state, If RMAppAttempt 
> receives the attempt_unregister event before attempt_kill event,  it'll 
> ignore the later attempt_kill event.  Hence, RMApp won't be able to move to 
> KILLED state and stays at KILLING state forever.

This message was sent by Atlassian JIRA

Reply via email to