[
https://issues.apache.org/jira/browse/YARN-6825?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16087177#comment-16087177
]
Rohith Sharma K S commented on YARN-6825:
-----------------------------------------
I think 80% of configured jute buffer should be taken implicitly rather than
allowing admin to configure 80% of jute buffer. This gives a buffered space for
allowing additional information into ApplicationStateData.
And also diagnosis error message length i.e 64Kb should be private
configurations. Otherwise, this issue will be there forever if there is any
misconfigured length of diagnosis errors.
> RM quit due to ApplicationStateData exceed the limit size of znode in zk
> ------------------------------------------------------------------------
>
> Key: YARN-6825
> URL: https://issues.apache.org/jira/browse/YARN-6825
> Project: Hadoop YARN
> Issue Type: Bug
> Components: resourcemanager
> Reporter: Rohith Sharma K S
>
> YARN-5006 fixes this issue by strict validation for ApplicationStateData
> length against 1MB(default max jute buffer) during application submission
> only. There is possibility of thrashing as dead zone was not properly
> defined/taken care.
> But it do not consider scenarios where ApplicationStateData can be increased
> later point of time i.e
> # If app is submitted with less than 1MB during submission, later updated
> like queue name or life time value or priority is changed. The app update
> call will be sent to statestore which cause same issue because
> ApplicationStateData length has increased.
> # Consider there is no app update, but final state are stored in ZK. This
> adds up several fields such finishTime, finalState, finalApplicationState.
> This increases size of ApplicationStateData.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
