[
https://issues.apache.org/jira/browse/FLINK-25377?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17480616#comment-17480616
]
Tamir Sagi edited comment on FLINK-25377 at 1/23/22, 12:22 PM:
---------------------------------------------------------------
Yes I see that in flink# org.apache.flink.kubernetes.utils.KubernetesUtils
resourceRequirementsBuilder
.addToRequests(configKey, resourceQuantity)
.addToLimits(configKey, resourceQuantity);
1. In general, kubernetes best practice regarding memory is request & limit
should be equal. in that way it's guaranteed to have enough memory when you
really need that.(for CPU is not that critical)
2. I checked your code,
[https://github.com/jeff-zou/flink/blob/xsj/flink-kubernetes/src/main/java/org/apache/flink/kubernetes/utils/KubernetesUtils.java#L350-L361]
in my opinion, a better solution is to have dedicated keys where the default is
request=limit. The user might/might not override them. it should always set it
accordingly.
was (Author: JIRAUSER283777):
Yes I see that in flink# org.apache.flink.kubernetes.utils.KubernetesUtils
resourceRequirementsBuilder
.addToRequests(configKey, resourceQuantity)
.addToLimits(configKey, resourceQuantity);
1. In general, kubernetes best practice regarding memory is to set an equal
value for request & limit. in that way it's guaranteed to have enough memory
when you really need that.(for CPU is not that critical)
2. I checked your code,
[https://github.com/jeff-zou/flink/blob/xsj/flink-kubernetes/src/main/java/org/apache/flink/kubernetes/utils/KubernetesUtils.java#L350-L361]
in my opinion, a better solution is to have dedicated keys where the default is
request=limit. The user might/might not override them. it should always set it
accordingly.
> kubernetes's request/limit resource can been seperated?
> -------------------------------------------------------
>
> Key: FLINK-25377
> URL: https://issues.apache.org/jira/browse/FLINK-25377
> Project: Flink
> Issue Type: New Feature
> Components: Deployment / Kubernetes
> Affects Versions: 1.13.2
> Environment: flink 1.13.2
> kubernetes
> Reporter: jeff-zou
> Priority: Major
>
> now my kubernetes cluster has very low CPU utilization,but cannt publish more
> tasks,because resources are already occupied by Pod request.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
