hadoop-mapreduce-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jason Lowe (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (MAPREDUCE-5670) CombineFileRecordReader should report progress when moving to the next file
Date Thu, 23 Jan 2014 23:48:45 GMT

    [ https://issues.apache.org/jira/browse/MAPREDUCE-5670?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13880521#comment-13880521

Jason Lowe commented on MAPREDUCE-5670:

Thanks for updating the patch, looking better.  Couple of things on the latest patch:

* We should not loop for the record reader next() to return false.  Since we explicitly set
up a bunch of empty files it should return false with only one invocation (we should assert
for this), and progress should still be called a total of 3 times.  If it somehow doesn't
return false in the first invocation then something is wrong, and continuing to loop is inviting
a hung test if for some reason it never does return false. ;-)
* outDir should be unique between the two tests or we risk having them collide if the tests
are run in parallel.  I suggest something like TestCombineFileRecordReader.class.getName()
which will be different between the two tests.
* Rather than loop to explicitly delete the individual files, we should just call FileUtil.fullyDelete
on outDir.

> CombineFileRecordReader should report progress when moving to the next file
> ---------------------------------------------------------------------------
>                 Key: MAPREDUCE-5670
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-5670
>             Project: Hadoop Map/Reduce
>          Issue Type: Bug
>          Components: mrv2
>    Affects Versions: 0.23.9
>            Reporter: Jason Lowe
>            Assignee: Chen He
>            Priority: Minor
>         Attachments: MR-5670.patch, MR-5670v2.patch
> If a combine split consists of many "empty" files (i.e.: no record found by the underlying
record reader) then theoretically a task can timeout due to lack of reported progress.

This message was sent by Atlassian JIRA

View raw message