[
https://issues.apache.org/jira/browse/YUNIKORN-3117?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18090155#comment-18090155
]
Peter Bacsko commented on YUNIKORN-3117:
----------------------------------------
I couple of days ago I did an analysis with AI on this subject to see how this
affects Yunikorn.
We have two options:
1) Use {{SharedDRAManager}} the way we use {{{}VolumeBinder{}}}. That is, check
how the plugin code calls it in various phases and do the same in the Yunikorn
codebase. I think this is not desirable, the DRA handling code in the
{{DyamicResources}} plugin is very complicated, duplicating that in any way is
very risky.
2) Retrieve plugin reference from the predicate manager during instantiation
and call Reserve/PreBind/Unreserve ourselves.
I definitely think #2 is the way to go. GPT-5.5 recommended a small lifecycle
wrapper layer around the plugin which handles the handoff between phases by
storing/passing {{CycleState}} around - a very reasonable idea.
> DRA support in YuniKorn (K8s 1.34 and later)
> --------------------------------------------
>
> Key: YUNIKORN-3117
> URL: https://issues.apache.org/jira/browse/YUNIKORN-3117
> Project: Apache YuniKorn
> Issue Type: New Feature
> Components: shim - kubernetes
> Reporter: Wilfred Spiegelenburg
> Assignee: Manikandan R
> Priority: Blocker
>
> DRA is partially going into GA for K8s 1.34. That means it is turned on.
> Investigate what the impact is on YuniKorn scheduling and if it can and
> should be supported.
> [Dynamic Resource Allocation
> (DRA)|https://kubernetes.io/docs/concepts/scheduling-eviction/dynamic-resource-allocation/]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
