[ 
https://issues.apache.org/jira/browse/FLINK-40010?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Martijn Visser resolved FLINK-40010.
------------------------------------
    Fix Version/s: 2.4.0
       Resolution: Fixed

Fixed in apache/flink:master ad2dee480a697d831d79ca1fd6cc501b198c7765

> RescaleTimelineITCase.testRescaleTerminatedByNoResourcesOrNoParallelismsChange
>  is flaky: requirements-update can miss the in-progress rescale
> ---------------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: FLINK-40010
>                 URL: https://issues.apache.org/jira/browse/FLINK-40010
>             Project: Flink
>          Issue Type: Bug
>          Components: Runtime / Coordination, Tests
>    Affects Versions: 2.4.0
>            Reporter: Martijn Visser
>            Assignee: Martijn Visser
>            Priority: Major
>             Fix For: 2.4.0
>
>
> testRescaleTerminatedByNoResourcesOrNoParallelismsChange fails on CI: the 
> awaited
> terminal reason NO_RESOURCES_OR_PARALLELISMS_CHANGE is never recorded, so the 
> wait times out (or, before https://issues.apache.org/jira/browse/FLINK-40009, 
> hangs).
> https://dev.azure.com/apache-flink/apache-flink/_build/results?buildId=76350&view=results
>  (leg: test_cron_azure core)
> https://dev.azure.com/apache-flink/apache-flink/_build/results?buildId=76242&view=results
>  (leg: test_cron_jdk11 core)
> Root cause: NO_RESOURCES_OR_PARALLELISMS_CHANGE is stamped by
> DefaultStateTransitionManager only on the rescale tracked when the manager
> (re-)enters its Idling phase. With the short shared cooldown, the cooldown 
> can elapse
> and the manager can reach Idling before the requirements-update RPC is 
> processed, so the UPDATE_REQUIREMENT rescale is created after Idling was 
> entered and never receives the terminal reason; it stays in-progress until 
> teardown cancels it (JOB_CANCELED).
> Fix: rebuild the fixture cluster with a cooldown (10s) that comfortably 
> outlasts the
> synchronous update RPC, so the update is processed in Cooldown and routed 
> back through Idling where the reason is stamped. Unlike 
> testRescaleTerminatedByResourceRequirementsUpdated (FLINK-39903), this case 
> must wait out the whole cooldown before the condition can be met, so the 
> cooldown is kept modest (10s) and the wait budget is widened to 60s.
> Related: FLINK-39902, FLINK-39903 (sibling races in the same class).



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

Reply via email to