[
https://issues.apache.org/jira/browse/YARN-8710?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16640361#comment-16640361
]
Billie Rinaldi commented on YARN-8710:
--------------------------------------
I think 3 retries are a bit low over an infinite failure validity interval.
Maybe 6 or 10 retries (with 1 retry every 30 seconds, that's 3 or 5 minutes of
retrying) over a failure validity interval of 10 or 15 minutes would be
appropriate. Components can configure different values if necessary.
The total retry count is controlled by
yarn.service.container-failure-per-component.threshold. It counts container
failures per component and does not count NM container retries (the AM is not
currently informed about retries, see YARN-8286). The default value is 10
failures per component over a 6 hour window.
> Service AM should set a finite limit on NM container max retries
> -----------------------------------------------------------------
>
> Key: YARN-8710
> URL: https://issues.apache.org/jira/browse/YARN-8710
> Project: Hadoop YARN
> Issue Type: Bug
> Components: yarn-native-services
> Reporter: Suma Shivaprasad
> Assignee: Suma Shivaprasad
> Priority: Major
> Attachments: YARN-8710.1.patch
>
>
> Container retries are currently set to a default of -1 in
> AbstractProviderService.buildContainerRetry. If this is not overridden via
> service spec with a finite value for yarn.service.container-failure.retry.max
> , this causes infinite NM reties for the container for ALWAYS/ON_FAILURE
> restart policy . Ideally it should try a finite number of time on the same NM
> and subsequently Service AM can retry on another node.
> We can set this to default value of 3.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
