[
https://issues.apache.org/jira/browse/HADOOP-4876?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12657276#action_12657276
]
Vivek Ratan commented on HADOOP-4876:
-------------------------------------
To summarize the problem:
HADOOP-4513 caused some code to be mistakenly removed from
TaskSchedulingMgr.updateQSIObject(). This code computed the number of pending
tasks, if required, from the list of waiting jobs. We need to know if there are
pending tasks in queue in order to start reclaiming capacity for the queue if
necessary. It's not enough to look at running jobs only. We must start the
timer for reclaiming capacity the moment a job is submitted to a queue that
does not have its Guaranteed Capacity. Consider the following case. You have a
queue that is empty, and its capacity is being utilized by other queues. Now
someone submits a job to it. That job will only show up in the run queue once
the initialization thread gets to it, so by looking at the jobs in the run
queue only, you're starting its reclaim timer later than you should.
> Capacity reclaimation for queue would not work as intended
> ----------------------------------------------------------
>
> Key: HADOOP-4876
> URL: https://issues.apache.org/jira/browse/HADOOP-4876
> Project: Hadoop Core
> Issue Type: Bug
> Components: contrib/capacity-sched
> Reporter: Sreekanth Ramakrishnan
> Assignee: Sreekanth Ramakrishnan
> Priority: Critical
> Attachments: HADOOP-4876-1.patch
>
>
> Capacity schedulers reclaim logic would not work as intended after
> [HADOOP-4513|https://issues.apache.org/jira/browse/HADOOP-4513] went in.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
