[
https://issues.apache.org/jira/browse/FLINK-9741?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17070980#comment-17070980
]
Biao Liu commented on FLINK-9741:
---------------------------------
Oops, sorry, I forgot this issue. I gave it a lower priority.
Anyway, we could discuss a bit here.
First, I think we'd better keep the metrics under {{JobManager}} due to
compatibility. [~trohrmann], [~chesnay], do you guys think it's acceptable that
forking the JVM metrics from {{JobManager}} to {{RM}}, {{Dispatcher}} or even
{{JobMaster}} ? We could do some optimization that they could share the same
metric instance. However it might still annoy user there are so many duplicated
metrics under different metric group, especially reported by {{MetricReporter}}.
> Register JVM metrics for each JM separately
> -------------------------------------------
>
> Key: FLINK-9741
> URL: https://issues.apache.org/jira/browse/FLINK-9741
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Coordination, Runtime / Metrics
> Affects Versions: 1.5.0, 1.6.0
> Reporter: Chesnay Schepler
> Assignee: Biao Liu
> Priority: Major
>
> Currently, the {{Dispatcher}} contains a {{JobManagerMetricGroup}} on which
> the JVM metrics are registered. the JobManagers only receive a
> {{JobManagerJobMetricGroup}} and don't register the JVM metrics.
> As the dispatcher and jobmanagers currently run in the same jvm, neither
> exposing their IDs to the metric system, this doesn't cause problem _right
> now_ as we can't differentiate between them anyway, but it will bite us down
> the line if either of the above assumptions is broken.
> For example, with the proposed exposure of JM/Dispatcher IDs in FLINK-9543 we
> would not expose JVM metrics tied to a JM, but only the Dispatcher.
> I propose to register all JVM metrics for each jobmanager separately to
> future proof the whole thing.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
