[ 
https://issues.apache.org/jira/browse/MAPREDUCE-3162?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13124758#comment-13124758
 ] 

Vinod Kumar Vavilapalli commented on MAPREDUCE-3162:
----------------------------------------------------

The deletion of app-dirs on each startLocalizer() is only happening in DCE 
case. And it seems like a copy-paste from the legacy code from TT. I don't see 
us deleting app-dirs like that in LCE case. The corner case mentioned there is 
also from the MRV1 TT world ("this will happen on a partial execution of 
localizeJob."), not sure if it even makes any sense now. Dropping that delete 
should do the trick for you at MAPREDUCE-3159?

Regarding the fork, in DCE, we are not even forking, we invoke the localizer 
directly. Your comment about avoiding unnecessary fork still applies to LCE 
though. But we cannot just avoid the fork directly without other related 
changes. One problem we've known from before is that since the app-dirs 
creation etc are done during the localization itself, not forking when there 
are no resources for a container would mean no app-dirs and so containers 
cannot even launch! *sigh*

We definitely need to move the creation of app-dirs etc out of 
resource-localization. That's a known issue but one which I delayed due to the 
bigger effort needed.
                
> Separate application-init and container-init event types in NM's 
> ApplicationImpl FSM
> ------------------------------------------------------------------------------------
>
>                 Key: MAPREDUCE-3162
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-3162
>             Project: Hadoop Map/Reduce
>          Issue Type: Improvement
>    Affects Versions: 0.23.0
>            Reporter: Todd Lipcon
>            Assignee: Todd Lipcon
>            Priority: Minor
>         Attachments: mr-3162.txt
>
>
> Currently, the ApplicationImpl receives an INIT_APPLICATION event on every 
> container initialization. Only on the first one does it really mean to init 
> the application, whereas all subsequent events are for specific containers. 
> This JIRA is to separate the events into INIT_APPLICATION, sent once and only 
> once per application, and INIT_CONTAINER, which is sent for every container. 
> The first container sends INIT_APPLICATION followed by INIT_CONTAINER.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: 
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to