[
https://issues.apache.org/jira/browse/YUNIKORN-689?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Chaoran Yu updated YUNIKORN-689:
--------------------------------
Description:
I observed that when my cluster was really busy and all incoming pods were
pending, after some time, YuniKorn still assigned a node for many (I forgot if
it was all) pending pods, even though kubelet thought that the node was out of
CPUs. Interestingly, all the pods that got an assignment were assigned to the
same node, even though my cluster had tens of nodes. Please see the attached
screenshot for what happened.
For context, all my pending pods were Spark pods and most of them were
placeholders. The rest were the real driver pods.
was:
I observed that when my cluster was really busy and all incoming pods were
pending, after some time, YuniKorn still assigned a node for many (I forgot if
it was all) pending pods, even though kubelet thought that the node was out of
CPUs. Please see the attached screenshot for what happened.
For context, all my pending pods were Spark pods and most of them were
placeholders. The rest were the real driver pods.
> Pods could be scheduled on nodes that don't have enough CPUs
> ------------------------------------------------------------
>
> Key: YUNIKORN-689
> URL: https://issues.apache.org/jira/browse/YUNIKORN-689
> Project: Apache YuniKorn
> Issue Type: Bug
> Components: shim - kubernetes
> Reporter: Chaoran Yu
> Priority: Major
> Attachments: Screen Shot 2021-05-22 at 11.33.30 AM.jpg
>
>
> I observed that when my cluster was really busy and all incoming pods were
> pending, after some time, YuniKorn still assigned a node for many (I forgot
> if it was all) pending pods, even though kubelet thought that the node was
> out of CPUs. Interestingly, all the pods that got an assignment were assigned
> to the same node, even though my cluster had tens of nodes. Please see the
> attached screenshot for what happened.
> For context, all my pending pods were Spark pods and most of them were
> placeholders. The rest were the real driver pods.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
