[
https://issues.apache.org/jira/browse/YUNIKORN-796?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Wilfred Spiegelenburg updated YUNIKORN-796:
-------------------------------------------
Target Version: 1.0.0
The node selection is based on the sorting policy. Bin packing would
concentrate pods on a limited set of nodes. Fair tries to distribute it evenly,
there is no guarantee.
There are however more things that do influence it: for instance (anti)
affinity settings on a pod. If a pod asks for disk mounts etc it could also
mean that a set of node cannot satisfy that request and will thus not be used.
You should be able to see those details. Those checks are handled as part of
the predicates etc. Do you have logs that show the grouping?
> Pod scheduling could disproportionately concentrate on one node
> ---------------------------------------------------------------
>
> Key: YUNIKORN-796
> URL: https://issues.apache.org/jira/browse/YUNIKORN-796
> Project: Apache YuniKorn
> Issue Type: Bug
> Components: core - scheduler
> Reporter: Chaoran Yu
> Priority: Major
>
> We've observed a few cases when YK keeps scheduling incoming pods to just one
> node, even though the cluster has many other nodes that have abundant
> resources. Not sure what could be causing this behavior because we don't see
> it in every K8s cluster we have. Maybe somewhere in the node sorting
> algorithm, a node could be favored over all other nodes under some conditions?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
