[
https://issues.apache.org/jira/browse/YARN-6895?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16109701#comment-16109701
]
Miklos Szegedi commented on YARN-6895:
--------------------------------------
Thank you, [~yufeigu].
We will do normal reservation, if there are no active preemptions on the node
for the app. Does this answer your question? There is still reservation on
other nodes if we preempt on one node but that should not be the cause of this
regression, since that logic has been around before YARN-6432.
{code}
// The desired container won't fit here, so reserve
// Reserve only, if not reserved for preempted resources, otherwise
// we may end up with duplicate reservations
if (isReservable(capability) &&
!node.isPreemptedForApp(this) &&
reserve(pendingAsk.getPerAllocationResource(), node, reservedContainer,
type, schedulerKey)) {
{code}
I had a patch with a single class implementation but it was rejected by the
reviewers. I think we can revisit but I would not add too many changes to this
Jira for simplicity.
> [FairScheduler] Preemption reservation may cause regular reservation leaks
> --------------------------------------------------------------------------
>
> Key: YARN-6895
> URL: https://issues.apache.org/jira/browse/YARN-6895
> Project: Hadoop YARN
> Issue Type: Bug
> Components: fairscheduler
> Affects Versions: 3.0.0-alpha4
> Reporter: Miklos Szegedi
> Assignee: Miklos Szegedi
> Priority: Blocker
> Attachments: YARN-6895.000.patch
>
>
> We found a limitation in the implementation of YARN-6432. If the container
> released is smaller than the preemption request, a node reservation is
> created that is never deleted.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
