[
https://issues.apache.org/jira/browse/FLINK-25910?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17489355#comment-17489355
]
Gabor Somogyi commented on FLINK-25910:
---------------------------------------
Propagating tokens during registration is either not enough or not good design
decision at all.
In 1.15 there is a feature where TaskManagers (TM for now on) can execute
workloads without being registered.
Here is an example scenario where workload can fail:
* TM1 has a valid token and is running some tasks.
* TM1 crashes
* TM2 is started to take over, and re-uses the working directory of TM1 (new
feature in 1.15!)
* TM2 recovers the previous slot allocations
* TM2 is informed about leading JM
* TM2 starts registration with RM
* TM2 offers slots to JobMaster
* TM2 accepts task submission from JobMaster
* ...some time later the registration completes...
All in all this must be considered when propagation is implemented.
> Propagate obtained delegation tokens to TaskManagers
> ----------------------------------------------------
>
> Key: FLINK-25910
> URL: https://issues.apache.org/jira/browse/FLINK-25910
> Project: Flink
> Issue Type: Sub-task
> Affects Versions: 1.15.0
> Reporter: Gabor Somogyi
> Priority: Major
>
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
