[
https://issues.apache.org/jira/browse/MAPREDUCE-682?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12732418#action_12732418
]
Hemanth Yamijala commented on MAPREDUCE-682:
--------------------------------------------
Looks fine. A couple of minor points:
- In FaultyTrackersInfo.blacklistTracker, the trackers we get for a hostname
are in a synchronized set and we are iterating over them. The javadoc of the
Collections.synchronizedSet says that when iterating we should guard it in a
synchronized block.
- As far as I can see, cancelAllReservations need not be synchronized now,
because all code paths seem to be coming after locking the jobtracker. Can you
please verify this once ? Possibly add a comment stating the assumption on
cancelAllReservations ?
- In the test case, can we also blacklist a different tracker (with
reservations) due to node health check reasons and make sure that both
trackers' reservations are removed. It just makes the test case cover more
scenarios with little work.
- Please add a comment for the cleanup steps in the test cases so its clear why
we need to do those steps.
> Reserved tasktrackers should be removed when a node is globally blacklisted
> ---------------------------------------------------------------------------
>
> Key: MAPREDUCE-682
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-682
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: jobtracker
> Affects Versions: 0.21.0
> Reporter: Hemanth Yamijala
> Attachments: mapreduce-682-1.patch
>
>
> When support was added to reserve tasktrackers for high RAM jobs per
> MAPREDUCE-516, we missed removing reservations on tasktrackers that are
> globally blacklisted. This is not a major concern, just that the reservation
> might cause the job to finish a little later than if the reservation is
> removed when the blacklisting happens.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
