Page MenuHomePhabricator

Upgrade flink-k8s-operator to 1.12.1 or downgrade to 1.11
Closed, ResolvedPublic

Description

We are running the 1.12.0 version of the flink-k8s-operator and it appears that this version is hit by a critical bug that will prevent jobs from upgrading (deploying new version): https://issues.apache.org/jira/browse/FLINK-38033

The bug appears fixed in 1.12.1 so we could upgrade to this or possibly downgrade to 1.11.

Event Timeline

dcausse renamed this task from Upgrade flink-k8s-operator to 1.12.1 or downgrade 1.11 to Upgrade flink-k8s-operator to 1.12.1 or downgrade to 1.11.Sep 5 2025, 3:29 PM
dcausse triaged this task as High priority.

Change #1185846 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/docker-images/production-images@master] flink-kubernetes-operator: upgrade to 1.12.1

https://gerrit.wikimedia.org/r/1185846

Change #1185846 merged by Brouberol:

[operations/docker-images/production-images@master] flink-kubernetes-operator: upgrade to 1.12.1

https://gerrit.wikimedia.org/r/1185846

~ ❯ docker pull docker-registry.discovery.wmnet/flink-kubernetes-operator:1.12.1-wmf0
1.12.1-wmf0: Pulling from flink-kubernetes-operator
653eb674aee2: Pull complete
ba36c88ab901: Downloading [========================>                          ]  37.14MB/74.71MB

The docker-registry.discovery.wmnet/flink-kubernetes-operator:1.12.1-wmf0 image was published and is now available.

Change #1185849 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in staging-codfw

https://gerrit.wikimedia.org/r/1185849

Change #1185850 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in staging-eqiad

https://gerrit.wikimedia.org/r/1185850

Change #1185851 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in codfw

https://gerrit.wikimedia.org/r/1185851

Change #1185852 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in eqiad

https://gerrit.wikimedia.org/r/1185852

Change #1185853 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in dse-k8s-eqiad

https://gerrit.wikimedia.org/r/1185853

Change #1185854 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1

https://gerrit.wikimedia.org/r/1185854

Change #1185850 merged by Brouberol:

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in staging-eqiad

https://gerrit.wikimedia.org/r/1185850

Done some testing :
reproduce the bug on cirrus-streaming-updater (done twice)

  1. destroy the job: helmfile destroy -e staging -i => OK
  2. deploy the job: helmfile deploy -e staging -i => OK
  3. restart the job: helmfile apply -e staging -i --selector name=producer --set app.restartNonce=2 => OK the job recovered

test suspend/resume

  1. suspend the job: helmfile apply -e staging -i --selector name=producer --set app.job.state=suspended => OK the job stopped, small caveat the jobmanager pod remained alive
  2. resume the job: helmfile apply -e staging -i --selector name=producer --set app.job.state=running => OK, the stale jobmanager pod disappeared and new one was created

test manual savepoint

  • take a savepoint: helmfile apply -e staging -i --selector name=producer --set app.job.savepointTriggerNonce=2 => OK, savepoint taken

test restarting another job without destroying the release first (rdf-streaming-updater)

  • restart the job: helmfile apply -e staging -i --selector name=wikidata --set app.restartNonce=2 => OK the job restarted on the right savepoint

@brouberol thanks, the bug is gone and all normal operations appear to work OK in staging.
A minor annoyance is that when we suspend the job flink does not appear to be willing to finish the jobmanager pod, it might stay idle. We rarely suspend for a long time so I believe that this is acceptable. I haven't done that in a while so it's very likely that it is an existing problem and probably not directly related to the operator but more to flink not willing to stop when it goes idle.

Change #1185849 merged by jenkins-bot:

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in staging-codfw

https://gerrit.wikimedia.org/r/1185849

Change #1185852 merged by jenkins-bot:

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in eqiad

https://gerrit.wikimedia.org/r/1185852

Change #1185851 merged by jenkins-bot:

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in codfw

https://gerrit.wikimedia.org/r/1185851

Change #1185853 merged by jenkins-bot:

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 in dse-k8s-eqiad

https://gerrit.wikimedia.org/r/1185853

Change #1185854 merged by jenkins-bot:

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1

https://gerrit.wikimedia.org/r/1185854

Change #1185872 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 (fix typo)

https://gerrit.wikimedia.org/r/1185872

Change #1185872 merged by Brouberol:

[operations/deployment-charts@master] flink-kubernetes-operator: upgrade to 1.12.1 (fix typo)

https://gerrit.wikimedia.org/r/1185872

All operators have been upgraded.

Tested one job in codfw and it worked well, thanks!

Change #1187416 had a related patch set uploaded (by Brouberol; author: Brouberol):

[operations/docker-images/production-images@master] flink: build flink 1.20 on top of bookworm/jdk 17

https://gerrit.wikimedia.org/r/1187416

Change #1187416 merged by Brouberol:

[operations/docker-images/production-images@master] flink: build flink 1.20 on top of bookworm/jdk 17

https://gerrit.wikimedia.org/r/1187416