[
https://issues.apache.org/jira/browse/MAPREDUCE-2429?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13048984#comment-13048984
]
Siddharth Seth commented on MAPREDUCE-2429:
-------------------------------------------
A little more context on the reason for the patch.
If a task is Killed (speculation or other reasons), there's a possibility that
the TaskTracker will end up losing track of the slot(s) used by the task. The
jobTracker will continue assigning tasks to the TT though.
The attemptId is re-used for cleanup tasks, there's a race where if there is a
delay in processing the SIGTERM sent to the child, the TT will allocate slots
for the CLEANUP task but will not actually launch it and release the slots.
(The old task sends in a done / commitPending which is applied to the TIP of
the cleanup task).
The patch validates the caller jvm - for all calls in TaskUmbilicalProtocol.
> Check jvmid during task status report
> -------------------------------------
>
> Key: MAPREDUCE-2429
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2429
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: tasktracker
> Affects Versions: 0.21.0, 0.22.0, 0.23.0
> Reporter: Arun C Murthy
> Assignee: Siddharth Seth
> Fix For: 0.20.204.0, 0.23.0
>
> Attachments: MR2429-1.patch, MR2429.patch, MR2429_trunk.patch
>
>
> Currently TT doens't check to ensure jvmid is relevant during communication
> with the Child via TaskUmbilicalProtocol.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
