[
https://issues.apache.org/jira/browse/FLINK-15641?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17079048#comment-17079048
]
Canbin Zheng commented on FLINK-15641:
--------------------------------------
Ah, #1 solves my original question. For #2 and #3, I think you are talking
about the Application mode, for the K8s per-job, internally we compiled the job
graph on the client-side, so we need to advertise and download it via the init
Container. Currently, I am not very sure whether we should also support
client-side per job mode on K8s yet.
Back to #1, I have another concern,
{quote}The schema could "file://", "hdfs://", "s3://", etc. If it is a remote
path, then we need to start the init container to download these jars and share
them with JM/TM container via volume.
{quote}
We'd better extend the schema to support local:// as well, local:// means path
in the image, while file:// means path on the client-sde.
> Support to start init container
> -------------------------------
>
> Key: FLINK-15641
> URL: https://issues.apache.org/jira/browse/FLINK-15641
> Project: Flink
> Issue Type: Sub-task
> Components: Deployment / Kubernetes
> Reporter: Yang Wang
> Priority: Major
>
> >> Why do we need init container?
> The init container could be used to prepare the use jars and dependencies.
> Then we could always set the user image to Flink official image both for
> standalone per-job on K8s or native K8s per-job. When the JobManager and
> TaskManager container launched, the user jars will already exist there. I
> think many users are running standalone per-job cluster in production by
> using this way.
> The init container only works for K8s cluster.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
