You are viewing a plain text version of this content. The canonical link for it is here.
Posted to mapreduce-issues@hadoop.apache.org by "Jason Lowe (JIRA)" <ji...@apache.org> on 2012/11/10 23:57:12 UTC
[jira] [Updated] (MAPREDUCE-4425) Speculation + Fetch failures can
lead to a hung job
[ https://issues.apache.org/jira/browse/MAPREDUCE-4425?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Jason Lowe updated MAPREDUCE-4425:
----------------------------------
Attachment: MAPREDUCE-4425.patch
This is similar to the KILL_WAIT hang reported in MAPREDUCE-4751 because it's caused by incorrect bookkeeping in TaskImpl. Since it ignores attempts completing after the task succeeds, it incorrectly thinks it has uncompleted attempts running after a fetch failure and therefore doesn't launch a new attempt.
Patch to fix the bookkeeping for attempts that complete while we're in the SUCCEEDED state so subsequent fetch failures will cause a new attempt to be launched.
> Speculation + Fetch failures can lead to a hung job
> ---------------------------------------------------
>
> Key: MAPREDUCE-4425
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-4425
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: mrv2
> Affects Versions: 0.23.1
> Reporter: Siddharth Seth
> Priority: Critical
> Attachments: MAPREDUCE-4425.patch
>
>
> After a task goes to SUCCEEDED, FAILED/KILLED attempts are ignored.
> 1. attemp1 starts
> 2. speculative attempt starts
> 3. attempt 1 completes - Task moves to SUCCEEDED state
> 4. speculative attempt is KILLED
> 5. T_ATTEMPT_KILLED is ignored.
> 6. attemp1 1 fails with TOO_MANY_FETCH_FAILURES
> The job will effectively hang, since a new task attempt isn't started.
--
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