[ 
https://issues.apache.org/jira/browse/YARN-1055?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13739040#comment-13739040
 ] 

Vinod Kumar Vavilapalli commented on YARN-1055:
-----------------------------------------------

This is a new issue with Hadoop 2 completely - we've added new failure 
conditions. All the apps handing AM restarts is really the right way forward 
given AMs can now run on random compute nodes that can just fail any time. 
Offline I started engaging some of Pig/Hive community folks. For MR, enough 
work is already done. Oozie needs to follow suit too.

Till work-preserving restart is finished, this is a real pain on RM restarts. 
Which is why I am proposing that oozie set max-attempts to 1 for its launcher 
action so that there are no split brain issues - RM restart or otherwise. Oozie 
has a retry mechanism anyways which will then be submitted as a new application.

Adding a separate knob just for restart is a hack I don't see any value of. If 
I read your proposal correctly, for launcher jobs, you will set 
restart.am.on.rm.restart to 1 and  restart.am.on.on.failure > 1. Right? Which 
is not correct as I repeated - node failures will cause the same split brain 
issues.
                
> Handle app recovery differently for AM failures and RM restart
> --------------------------------------------------------------
>
>                 Key: YARN-1055
>                 URL: https://issues.apache.org/jira/browse/YARN-1055
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: resourcemanager
>    Affects Versions: 2.1.0-beta
>            Reporter: Karthik Kambatla
>
> Ideally, we would like to tolerate container, AM, RM failures. App recovery 
> for AM and RM currently relies on the max-attempts config; tolerating AM 
> failures requires it to be > 1 and tolerating RM failure/restart requires it 
> to be = 1.
> We should handle these two differently, with two separate configs.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to