[
https://issues.apache.org/jira/browse/FLINK-21128?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17272058#comment-17272058
]
Yang Wang commented on FLINK-21128:
-----------------------------------
[~chesnay] Thanks for creating this issue. I think your suggestion makes a lot
of sense to me. Currently, the native K8s integration implementation is in a
hybrid state, which uses both the JAVA codes and scripts to launch the
JobManager and TaskManager pods. It is not very easy to maintain.
I think we could introduce some dedicated scripts(e.g.
{{kubernetes-application.sh}}, {{kubernetes-jobmanager.sh}},
{{kubernetes-taskmanager.sh}}) for native Kubernetes. Theses scripts will help
with constructing the classpath, setting the logging and etc. Of cause, the
duplicated logics implemented by java in {{flink-kubernetes}} module will be
removed.
Apart from these scripts, Flink java codes will generate some dynamic config
options(e.g. {{-D jobmanager.memory.off-heap.size=134217728b -D
jobmanager.memory.jvm-overhead.min=858993472b}}) and easily pass them to the
script.
> Rethink how Kubernetes interfaces with our docker image
> -------------------------------------------------------
>
> Key: FLINK-21128
> URL: https://issues.apache.org/jira/browse/FLINK-21128
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / Kubernetes, Deployment / Scripts
> Reporter: Chesnay Schepler
> Priority: Major
> Fix For: 1.13.0
>
>
> Our native Kubernetes integration assembles a {{java}} command that is passed
> to the docker-script, which first does some magical stuff to create the
> classpath
> {code}
> # Set the Flink related environments
> export _FLINK_HOME_DETERMINED=true
> . $FLINK_HOME/bin/config.sh
> export FLINK_CLASSPATH="`constructFlinkClassPath`:$INTERNAL_HADOOP_CLASSPATHS"
> {code}
> and then evaluates the given java command.
> This magical stuff is a bit problematic for a few reasons:
> 1) it is _only_ required for Kubernetes to work; it should not be executed in
> any other case
> 2) it appears to be required for running anything in Flink that is not part
> of the docker-entrypoint API (e.g., jobmanager) when it isn't.
> 3) it unnecessarily spreads kubernetes-specific logic across the flink and
> flink-docker repositories.
> A simple change would be to introduce dedicated Kubernetes scripts into the
> distribution, which for starters will just import {{config.sh}}, export
> FLINK_CLASSPATH and evaluate the given command.
> In the long-run it would be good if we would not create a java command but
> instead relied on the existing scripts which do exactly that already.
> As is stands we're duplicating code across the distribution and the
> Kubernetes module, which already caused troubles when we did the log4j2
> migration.
> Ideally we find a way to reuse the existing code in the Kubernetes module for
> generating these commands in the existing scripts, e.g., by moving them to
> the BashJavaUtils.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
