[
https://issues.apache.org/jira/browse/MAPREDUCE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Siddharth Seth updated MAPREDUCE-3721:
--------------------------------------
Attachment: MR-3721.txt
Patch adds another variable - commitMemory (Fetch complete size). A merge is
triggered only if this size exceeds mergeThreshold. Added a check to ensure
mergeThreshold is greater than the maxSingleShuffleLimit.
Earlier - usedMemory (reserved) was used for this computation - which meant a
single segment way below mergeThreshold could lead to a merge to disk.
Have run several gridmix runs with the patch applied - without a hang. Not
including a unit test - writing one would likely change way more in the shuffle
code to be able to recreate the scenario.
> Race in shuffle can cause it to hang
> ------------------------------------
>
> Key: MAPREDUCE-3721
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-3721
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: mrv2
> Affects Versions: 0.23.0
> Reporter: Siddharth Seth
> Assignee: Siddharth Seth
> Priority: Blocker
> Attachments: MR-3721.txt
>
>
> If all current {{Fetcher}}s complete while an in-memory merge is in progress
> - shuffle could hang.
> Specifically - if the memory freed by an in-memory merge does not bring
> {{MergeManager.usedMemory}} below {{MergeManager.memoryLimit}} and all
> current Fetchers complete before the in-memory merge completes, another
> in-memory merge will not be triggered - and shuffle will hang. (All new
> fetchers are asked to WAIT).
--
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
