[
https://issues.apache.org/jira/browse/FLINK-4079?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15332032#comment-15332032
]
Maximilian Michels commented on FLINK-4079:
-------------------------------------------
Exactly, the yarn properties file is a means to easily submit jobs against a
long-running Flink cluster (so called yarn session). Actually, it is the only
proper way at the moment besides figuring out the JobManager location manually
and essentially treating the Yarn cluster as a Standalone cluster.
I would also prefer to get rid of the properties file and only use the yarn
application id instead. We could probably do that for one of the next releases
but it would be a breaking change for the 1.1 release.
> YARN properties file used for per-job cluster
> ---------------------------------------------
>
> Key: FLINK-4079
> URL: https://issues.apache.org/jira/browse/FLINK-4079
> Project: Flink
> Issue Type: Bug
> Affects Versions: 1.0.3
> Reporter: Ufuk Celebi
> Fix For: 1.0.4
>
>
> YARN per job clusters (flink run -m yarn-cluster) rely on the hidden YARN
> properties file, which defines the container configuration. This can lead to
> unexpected behaviour, because the per-job-cluster configuration is merged
> with the YARN properties file (or used as only configuration source).
> A user ran into this as follows:
> - Create a long-lived YARN session with HA (creates a hidden YARN properties
> file)
> - Submits standalone batch jobs with a per job cluster (flink run -m
> yarn-cluster). The batch jobs get submitted to the long lived HA cluster,
> because of the properties file.
> [~mxm] Am I correct in assuming that this is only relevant for the 1.0 branch
> and will be fixed with the client refactoring you are working on?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
