[
https://issues.apache.org/jira/browse/YARN-11965?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18089188#comment-18089188
]
ASF GitHub Bot commented on YARN-11965:
---------------------------------------
K0K0V0K commented on PR #8547:
URL: https://github.com/apache/hadoop/pull/8547#issuecomment-4709807027
Thanks @zhengchenyu for the change!
Overall, this looks good to me.
One thing that came to mind: if I understand correctly, the
`/ws/v1/cluster/metrics` endpoint currently has a fairly predictable response
size and response time. With this change, both could become dependent on the
number of partitions in the cluster.
As far as I know, a single ResourceManager can handle around 10,000
NodeManagers. In a hypothetical worst-case scenario where each NodeManager
belongs to a different partition, could you measure how the response time is
affected?
I’m not familiar with your use case, but have you considered exposing this
information through a separate endpoint instead? For example, something like
`/ws/v1/cluster/partitions/metrics` or` /ws/v1/cluster/partition-metrics` might
avoid introducing any regression in the response size or latency of the
existing /ws/v1/cluster/metrics endpoint while still making the additional data
available.
> Support partition-aware resource metrics in RM cluster metrics REST API
> -----------------------------------------------------------------------
>
> Key: YARN-11965
> URL: https://issues.apache.org/jira/browse/YARN-11965
> Project: Hadoop YARN
> Issue Type: Improvement
> Reporter: Chenyu Zheng
> Assignee: Chenyu Zheng
> Priority: Major
> Labels: pull-request-available
>
> When node labels are enabled, different labels represent separate resource
> pools. However, the RM REST API `/ws/v1/cluster/metrics` currently exposes
> fields such as totalMB and totalVirtualCores based on the default partition
> only. As a result, resources from non-default partitions are not visible to
> external resource management systems, which may incorrectly determine that
> the cluster has no available capacity after new labels are added.
> The API should expose cluster resource metrics across all partitions and
> provide partition-level metrics so clients can distinguish capacity and usage
> by node label.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
