[
https://issues.apache.org/jira/browse/FLINK-1950?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14516694#comment-14516694
]
Stephan Ewen commented on FLINK-1950:
-------------------------------------
Given that our non-heap overhead has a minimal constant component as well (2 *
netty direct buffers, a minimum of 256m perm gen space), I think having a
minimum cutoff of 380 mb seems quite reasonable for flink as well.
After that, it should be proportional, as the native pool, the stack space, etc
are all relative to the heap, if I recall correctly.
10% seems almost a bit too little, my gut feeling would have been 15% or so,
but if they ran a lot of experiments for that, it is probably a reasonable
number.
> Increase default heap cutoff ratio from 20% to 30% and move default value to
> ConfigConstants
> --------------------------------------------------------------------------------------------
>
> Key: FLINK-1950
> URL: https://issues.apache.org/jira/browse/FLINK-1950
> Project: Flink
> Issue Type: Task
> Components: YARN Client
> Affects Versions: 0.9
> Reporter: Robert Metzger
> Assignee: Robert Metzger
> Priority: Blocker
>
> It seems that too many production users are facing issues with YARN killing
> containers due to resource overusage.
> We can mitigate the issue by using only 70% of the specified memory for the
> heap.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
