[
https://issues.apache.org/jira/browse/MAPREDUCE-2480?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13031281#comment-13031281
]
Luke Lu commented on MAPREDUCE-2480:
------------------------------------
bq. To further clarify, the dependency, ideally, should not be on the
implementation but rather on some kind of interface exposed by the shuffle
service.
That's exactly my point. It should not depend on a specific implementation
class but rather an SPI that both the app and the shuffle implementation depend
on. In this case, the SPI is the service id at the moment. If you rename the
ShuffleHandler to ShuffleHandlerImpl the code would look wrong too.
Now you have a runtime classpath dependency on a specific shuffle
implementation for mr app. If you replace the jar with a different class and
update the config, the NM will happily load the service but the mr app won't
even run due to class not found exception.
> MR-279: mr app should not depend on hard-coded version of shuffle
> -----------------------------------------------------------------
>
> Key: MAPREDUCE-2480
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2480
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: mrv2
> Reporter: Luke Lu
> Assignee: Luke Lu
> Fix For: 0.23.0
>
> Attachments: mr-2480-shuffle-version-v1.patch
>
>
> The following commit introduced a dependency of shuffle with hard-coded
> version for mr app:
> {noformat}
> commit 6f69742140516be7493c9a9177b81d0516cc9539
> Author: Vinod Kumar Vavilapalli <[email protected]>
> Date: Wed May 4 06:53:52 2011 +0000
> Adding user log handling for YARN. Making NM put the user-logs on DFS and
> providing log-dump tools. Contributed by Vinod Kumar Vavilapalli.
> {noformat}
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
