[
https://issues.apache.org/jira/browse/HADOOP-4988?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12663970#action_12663970
]
Hemanth Yamijala commented on HADOOP-4988:
------------------------------------------
bq. removed code in assignTasks() that depended on queues with gc=0 being at
the end of the collection.
Vivek, started looking at this patch. If a queue has no capacity, we should not
be giving a task. The code removed in the patch would hand out a task to it,
which is wrong. What the fix should be is that previously, because of the sort
order of queues, since queues with 0 capacity came at the end, we assumed
there's no need to look at other queues. This we should change and start
looking at other queues as well. Makes sense ?
> An earlier fix, for HADOOP-4373, results in a problem with reclaiming
> capacity when one or more queues have a capacity equal to zero
> ------------------------------------------------------------------------------------------------------------------------------------
>
> Key: HADOOP-4988
> URL: https://issues.apache.org/jira/browse/HADOOP-4988
> Project: Hadoop Core
> Issue Type: Bug
> Components: contrib/capacity-sched
> Reporter: Vivek Ratan
> Priority: Blocker
> Attachments: 4988.1.patch
>
>
> HADOOP-4373 introduced a fix for queues with guaranteed capacity (gc) equal
> to zero. Part of the fix was in the queue comparator used to sort queues.
> Queues with gc=0 were placed at the end. This causes a problem with the code
> for reclaiming capacity, which assumes that queues are sorted based on free
> space available and that a queue with gc=0 is no different than a queue which
> is running at capacity. Because of this, the following problem can arise: if
> we have a system with at least one queue whose gc=0, we may fail to reclaim
> capacity for some queues.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
