[
https://issues.apache.org/jira/browse/FLINK-40558?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
ASF GitHub Bot updated FLINK-40558:
-----------------------------------
Labels: pull-request-available (was: )
> Apply in-place scaling only on SCALE spec diff type
> ---------------------------------------------------
>
> Key: FLINK-40558
> URL: https://issues.apache.org/jira/browse/FLINK-40558
> Project: Flink
> Issue Type: Improvement
> Components: Autoscaler, Kubernetes Operator
> Reporter: Dennis-Mircea Ciupitu
> Priority: Major
> Labels: pull-request-available
>
> Currently, the reconciler’s in-place scaling can be reached by a non-SCALE
> spec diff type. This case can only be met when the spec diff type is IGNORE,
> and the reconciliation state is UPGRADING.
> In reality this state is met when the status was patched and the operator
> crashed before performing the deploy (for first deployment cases) or
> restoring the job/cluster, by having at that time the deployment not deployed
> at all or the job/cluster fully cancelled/deleted. For a job this re-enters
> as an UPGRADE diff though, since its state is recorded as SUSPENDED before
> deploy, so it is already excluded, and from these crash cases only the
> session cluster stays on IGNORE.
> Luckily, today, as mostly everything happens synchronously, the scale method
> for both native and standalone both incidentally happen to return false in
> this kind of case, because:
> - In native mode, a session cluster has no job spec, so the in-place scaling
> support check returns false and the scale method will not return a success.
> - In standalone mode, there will be no TM deployment that is found and the
> scale method will not return a success.
> Still, there is a current gap is for the stateless suspension which happen
> async on Flink cluster side, even tho the ack was send back sync to the
> operator, and this can create a discrepancy by still keeping the job in a non
> terminal state that can pass the scale checks on the next reconciliation
> pass, depending on how fast this might happen.
> The above scenario is more like an edge one, but still, by keeping the scale
> logic called only for the SCALE diff type will keep both the code and logic
> clean and will help us avoiding unnecessary situations.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
