[ 
https://issues.apache.org/jira/browse/YUNIKORN-3332?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Ompal Singh updated YUNIKORN-3332:
----------------------------------
    Description: 
On restart we should not create placeholders that are already timed-out. In 
certain restart scenarios we could be creating a placeholder that is already 
timed out.
 * Stop YuniKorn

 * Create a pod with gang definition and a short placeholder timeout

 * Start YuniKorn

If the time between the pod creation and the start of YuniKorn is longer than 
the timeout the placeholders will immediately timeout on allocation.

We have two choices:
 * not create the placeholders and directly push the real pod

 * set the time out based on the restart time of YuniKorn

We must only do that for newly created placeholders not for existing ones.

  was:
The REST object for the queue does not expose the new attributes for the queues 
for ask back off and quota preemption.

 

We should expose these 4 in the REST object:
 * quotaPreemptionStartTime time.Time
 * isQuotaPreemptionRunning bool
 * unschedAskBackoff        uint64
 * askBackoffDelay          time.Duration


> Prevent timed out Placeholder creation
> --------------------------------------
>
>                 Key: YUNIKORN-3332
>                 URL: https://issues.apache.org/jira/browse/YUNIKORN-3332
>             Project: Apache YuniKorn
>          Issue Type: Task
>          Components: core - common
>            Reporter: Ompal Singh
>            Assignee: Ompal Singh
>            Priority: Major
>              Labels: newbie, pull-request-available
>
> On restart we should not create placeholders that are already timed-out. In 
> certain restart scenarios we could be creating a placeholder that is already 
> timed out.
>  * Stop YuniKorn
>  * Create a pod with gang definition and a short placeholder timeout
>  * Start YuniKorn
> If the time between the pod creation and the start of YuniKorn is longer than 
> the timeout the placeholders will immediately timeout on allocation.
> We have two choices:
>  * not create the placeholders and directly push the real pod
>  * set the time out based on the restart time of YuniKorn
> We must only do that for newly created placeholders not for existing ones.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to