[
https://issues.apache.org/jira/browse/FLINK-15288?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16998793#comment-16998793
]
Xintong Song commented on FLINK-15288:
--------------------------------------
[~azagrebin],
I agree with [~fly_in_gis] that this issue is indeed a blocker for 1.10.
We have the container cutoff for yarn / mesos for good reasons. Without the
cutoff, the container can easily use memory more than requested, and could get
killed by yarn / mesos for that. The problem is the same for k8s.
Moreover, the cutoff ratio is a common config option for all containerized
setups. The k8s integration, as a new containerized setup, should comply to
this common configuration. Otherwise, the users will need to understand that
there's a different JM memory configuration logic for k8s compared to other
setups.
If later we decided to change the JM memory configuration to remove the cutoff,
the users only need to know that cutoff is replaced by something else (e.g.,
jvm overhead), they do not need to understand how that change is different for
individual deployments.
> Starting jobmanager pod should respect containerized.heap-cutoff
> ----------------------------------------------------------------
>
> Key: FLINK-15288
> URL: https://issues.apache.org/jira/browse/FLINK-15288
> Project: Flink
> Issue Type: Sub-task
> Components: Deployment / Kubernetes
> Reporter: Yang Wang
> Assignee: Yang Wang
> Priority: Blocker
> Labels: pull-request-available
> Fix For: 1.10.0
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Starting jobmanager pod should respect containerized.heap-cutoff. The cutoff
> will be used to leave some memory for jvm off-heap, for example meta space,
> thread native memory and etc.
>
> {code:java}
> public static final ConfigOption<Float> CONTAINERIZED_HEAP_CUTOFF_RATIO =
> ConfigOptions
> .key("containerized.heap-cutoff-ratio")
> .defaultValue(0.25f)
> .withDeprecatedKeys("yarn.heap-cutoff-ratio")
> .withDescription("Percentage of heap space to remove from containers (YARN /
> Mesos / Kubernetes), to compensate" +
> " for other JVM memory usage.");
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
