[
https://issues.apache.org/jira/browse/YUNIKORN-3320?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18100281#comment-18100281
]
KAI-JAY TU commented on YUNIKORN-3320:
--------------------------------------
The fix from YUNIKORN-3318 (commit 3dbbb5d) fully addresses this issue. The
guard {{ask.GetRequiredNode() == ""}} at application.go:1430 prevents required
node reservations from being cancelled when the reservationWaitTimeout expires.
I've submitted a PR to add dedicated test coverage for this behavior:
- TestTryReservedAllocateRequiredNodeNotCancelledOnTimeout verifies that a
required node reservation is NOT cancelled even after 90 minutes (well beyond
the 60-minute timeout), contrasting with the existing test that shows normal
reservations ARE cancelled.
The preemption path in preemption.go is also already protected (line 172: {{if
res.alloc.requiredNode != "" ... continue}}).
PR: https://github.com/apache/yunikorn-core/pull/1113
> tryReservedAllocate can release required node reservations
> ----------------------------------------------------------
>
> Key: YUNIKORN-3320
> URL: https://issues.apache.org/jira/browse/YUNIKORN-3320
> Project: Apache YuniKorn
> Issue Type: Bug
> Components: core - scheduler
> Affects Versions: 1.9.0
> Reporter: Wilfred Spiegelenburg
> Priority: Major
> Labels: newbie, pull-request-available
>
> {{tryReservedAllocate}} cleans up reservations that should not be on the node
> any more: reservation is older than the {{reservationWaitTimeout}}. This does
> not take into account that there might be other required node reservations on
> that same node.
> Required node reservations should not be released from a node even if they
> are older than the timeout
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
