[
https://issues.apache.org/jira/browse/YUNIKORN-2665?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Manikandan R updated YUNIKORN-2665:
-----------------------------------
Description:
Gang app choose the first pod (who created the app) as originator pod which
becomes the real driver pod later. While processing gang app specifically after
the placeholder creation and in the process of replacement, restart can lead to
the below described incorrect behaviour:
During restore, there is no guarantee on the ordering of pods coming from K8s
lister especially when all the pods created with the same second timestamp. k8s
use the seconds based timestamp, which means all pods created with in same
second has same timestamp. During this situation, whichever pod comes first
from lister, YK designate it as originator pod. So, any placeholder could
become the originator pod and actual originator pod has been lost. This change
could cause rippling effects leading to weird behaviour and needs to be fixed.
was:
Gang app choose the first pod (who created the app) as originator pod which
becomes the real driver pod later. While processing gang app specifically after
the placeholder creation and in the process of replacement, restart can lead to
the below described incorrect behaviour:
During restore, there is no guarantee on the ordering of pods from K8s lister
especially when all the pods created with the same second timestamp. k8s use
the seconds based timestamp, which means all pods created with in same second
has same timestamp. During this situation, which pod comes first from lister,
YK designate it as originator pod. So, any placeholder could become the
originator pod and actual originator pod has lost. This change could cause
rippling effects and needs to be fixed.
> Gang app originator pod changes after restart
> ---------------------------------------------
>
> Key: YUNIKORN-2665
> URL: https://issues.apache.org/jira/browse/YUNIKORN-2665
> Project: Apache YuniKorn
> Issue Type: Bug
> Components: shim - kubernetes
> Affects Versions: 1.3.0, 1.4.0, 1.5.0, 1.5.1, 1.5.2
> Reporter: Manikandan R
> Assignee: Manikandan R
> Priority: Major
>
> Gang app choose the first pod (who created the app) as originator pod which
> becomes the real driver pod later. While processing gang app specifically
> after the placeholder creation and in the process of replacement, restart can
> lead to the below described incorrect behaviour:
> During restore, there is no guarantee on the ordering of pods coming from K8s
> lister especially when all the pods created with the same second timestamp.
> k8s use the seconds based timestamp, which means all pods created with in
> same second has same timestamp. During this situation, whichever pod comes
> first from lister, YK designate it as originator pod. So, any placeholder
> could become the originator pod and actual originator pod has been lost. This
> change could cause rippling effects leading to weird behaviour and needs to
> be fixed.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
