[
https://issues.apache.org/jira/browse/FLINK-32829?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ivan Stoiev updated FLINK-32829:
--------------------------------
Description:
In order to optimize kubernetes nodes CPU usage before operator, we did usually
set higher CPU limit than request at kubernetes level.
This configuration allowed a faster rollout/restart of job/task managers
because PODs can use unused CPU from node until they start processing.
In operator, the only possible way of setting CPU request/limit is under
job/taskmanager.resource.cpu, this sets kubernetes
pod.spec.resource.requests.cpu equals to pod.spec.resource.limits.cpu, and
there is no way to configure those settings differently.
Here is a ilustration of CPU usage of PODs (managers) as percent of CPU limits
under a rollout, using limits ≠ requests:
!image-2023-08-10-15-18-22-705.png!
Is this by design, or a possible missing feature?
was:
In order to optimize kubernetes nodes CPU usage before operator, we did usually
set higher CPU limit than request at kubernetes level.
This configuration allowed a faster rollout/restart of job/task managers
because PODs can use unused CPU from node until they start processing.
In operator, the only possible way of setting CPU request/limit is under
job/taskmanager.resource.cpu, this sets kubernetes
pod.spec.resource.requests.cpu equals to pod.spec.resource.limits.cpu, and
there is no way to configure those settings differently.
Here is a ilustration of CPU usage of PODs (managers) as percent of CPU limits
under a rollout:
!image-2023-08-10-15-18-22-705.png!
Is this by design, or a possible missing feature?
> Allow setting different values for CPU requests and limits
> ----------------------------------------------------------
>
> Key: FLINK-32829
> URL: https://issues.apache.org/jira/browse/FLINK-32829
> Project: Flink
> Issue Type: New Feature
> Components: Kubernetes Operator
> Reporter: Ivan Stoiev
> Priority: Minor
> Attachments: image-2023-08-10-15-18-22-705.png
>
>
> In order to optimize kubernetes nodes CPU usage before operator, we did
> usually set higher CPU limit than request at kubernetes level.
> This configuration allowed a faster rollout/restart of job/task managers
> because PODs can use unused CPU from node until they start processing.
> In operator, the only possible way of setting CPU request/limit is under
> job/taskmanager.resource.cpu, this sets kubernetes
> pod.spec.resource.requests.cpu equals to pod.spec.resource.limits.cpu, and
> there is no way to configure those settings differently.
> Here is a ilustration of CPU usage of PODs (managers) as percent of CPU
> limits under a rollout, using limits ≠ requests:
> !image-2023-08-10-15-18-22-705.png!
>
> Is this by design, or a possible missing feature?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
