[jira] [Commented] (FLINK-7087) Implement Flip-6 container entry point

["Eron Wright (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-7087?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16093723#comment-16093723 ] 

Eron Wright  commented on FLINK-7087:
-------------------------------------


What is the proposed design for embedding the job into the container image?  One of the key questions is whether the program's main method would be executed by the entrypoint, to obtain a job graph, or whether the job graph itself is stored in the image.   The latter has cleaner recovery semantics but implies a new variant of `flink run` that exports the plan rather than submitting it.    We must also answer how the entrypoint locates the program or job graph.

Can we describe how this would all work in K8?   In terms of controllers, pods etc.  Is the JM pod defined differently from the TM pod(s)?   There are pros/cons to be discussed, e.g. having a homogeneous set of pods where anyone could be elected as the leader, versus the fact that the JM has different memory requirements than does the TM.  As a stretch goal, consider how job upgrade will work in terms of K8 functionality.

Maybe an umbrella ticket should be created to track the K8 issues as a group.  

> Implement Flip-6 container entry point
> --------------------------------------
>
>                 Key: FLINK-7087
>                 URL: https://issues.apache.org/jira/browse/FLINK-7087
>             Project: Flink
>          Issue Type: Bug
>          Components: Cluster Management
>            Reporter: Till Rohrmann
>              Labels: flip-6
>
> In order to support Docker and K8 we have to provide a container entry point. In a first version, the container entry point could be similar to the standalone session mode just with the difference that we don't submit a job to the cluster. The job has to be contained in the container image or being at least retrievable by the entry point. In that sense the container entry point acts like a per-job standalone mode.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)