[ 
https://issues.apache.org/jira/browse/FLINK-34589?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18129874#comment-18129874
 ] 

Martijn Visser commented on FLINK-34589:
----------------------------------------

[~xtsong] [~mapohl] picking this up with FLINK-41059. The check already runs 
under {{PekkoRpcActor#handleRunAsync}}, which logs and continues, so production 
has been option 2 since 2021; what throws here synchronously is a bug or an 
invariant violation, backend failures surface asynchronously in the 
ActiveResourceManager. The case for option 1 is the {{checkState}} in 
{{ActiveResourceManager#checkResourceDeclarations}}: a counter drift would be 
logged on every declaration while the cluster stays under-provisioned. The case 
for option 2 is that a strategy bug would crash-loop the whole session cluster 
instead of leaving one job without slots. I lean to option 2 made explicit, one 
ERROR line naming the check and fatal errors still propagating. Does either of 
you still prefer option 1?

> FineGrainedSlotManager doesn't handle errors in the resource reconcilliation 
> step
> ---------------------------------------------------------------------------------
>
>                 Key: FLINK-34589
>                 URL: https://issues.apache.org/jira/browse/FLINK-34589
>             Project: Flink
>          Issue Type: Technical Debt
>          Components: Runtime / Coordination
>    Affects Versions: 1.19.0, 1.18.1, 1.20.0
>            Reporter: Matthias Pohl
>            Priority: Minor
>
> I noticed during my work on FLINK-34427 that the reconcilliation is scheduled 
> periodically when starting the {{SlotManager}}. But it doesn't handle errors 
> in this step. I see two options here:
> 1. Fail fatally because such an error might indicate a major issue with the 
> RM backend.
> 2. Log the failure and continue the scheduled task even in case of an error.
> My understanding is that we're just not able to recreate TaskManagers which 
> should be a transient issue and could be resolved in the backend (YARN, k8s). 
> That's why I would lean towards option 2.
> [~xtsong] WDYT?



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to