Uploaded image for project: 'Flink'
  1. Flink
  2. FLINK-34451

[Kubernetes Operator] Job with restarting TaskManagers uses wrong/misleading fallback approach



    • Bug
    • Status: Open
    • Major
    • Resolution: Unresolved
    • kubernetes-operator-1.6.1
    • None
    • Kubernetes Operator
    • None
    • Operator version: 1.7.1

      Flink version 1.18.0

      HA JobManagers

      Adaptive scheduler mode using the operator's autoscaler

      Checkpointing at an interval of 60s

      Upgrade mode savepoint



      We had a situation where TaskManagers were constantly restarting from OOM. We're using the Adaptive scheduler with the Kubernetes Operator, and a restart strategy of exponential backoff, and so the JobManagers remained alive. We're also using savepoint upgrade mode. 

      When we tried to remedy the situation by raising the direct memory allocation to the pods, we were surprised that Flink used the last savepoint taken, rather than the checkpoint. This was unfortunate for us because we are on adaptive scheduler and the job hasn't changed in some time, so this last savepoint was 6 days old! Meanwhile, checkpoints were taken every minute up until failure. I can confirm the HA metadata existed in the configmaps, and the corresponding checkpoints existed in remote storage for it to access. Plus, no Flink version changes were in the deployment.

      The Operator logs reported that it was using last-state recovery in this situation:

      2024-02-15 19:38:38,252 o.a.f.k.o.l.AuditUtils         [INFO ][job-name] >>> Event  | Info    | SPECCHANGED     | UPGRADE change(s) detected (Diff: FlinkDeploymentSpec[image : image:0a7c41b -> image:ebebc53, restartNonce : null -> 100]), starting reconciliation.
      2024-02-15 19:38:38,252 o.a.f.k.o.r.d.AbstractJobReconciler [INFO ][job-name] Upgrading/Restarting running job, suspending first...
      2024-02-15 19:38:38,260 o.a.f.k.o.r.d.ApplicationReconciler [INFO ][job-name] Job is not running but HA metadata is available for last state restore, ready for upgrade
      2024-02-15 19:38:38,270 o.a.f.k.o.l.AuditUtils         [INFO ][job-name] >>> Event  | Info    | SUSPENDED       | Suspending existing deployment.
      2024-02-15 19:38:38,270 o.a.f.k.o.s.NativeFlinkService [INFO ][job-name] Deleting JobManager deployment while preserving HA metadata.	
      2024-02-15 19:38:40,431 o.a.f.k.o.l.AuditUtils         [INFO ][job-name] >>> Status | Info    | UPGRADING       | The resource is being upgraded 
      2024-02-15 19:38:40,532 o.a.f.k.o.l.AuditUtils         [INFO ][job-name] >>> Event  | Info    | SUBMIT          | Starting deployment
      2024-02-15 19:38:40,532 o.a.f.k.o.s.AbstractFlinkService [INFO ][job-name] Deploying application cluster requiring last-state from HA metadata
      2024-02-15 19:38:40,538 o.a.f.k.o.u.FlinkUtils         [INFO ][job-name] Job graph in ConfigMap job-name-cluster-config-map is deleted 

      But when the job booted up, it reported restoring from savepoint:

      2024-02-15 19:39:03,887 INFO  org.apache.flink.runtime.checkpoint.CheckpointCoordinator    [] - Restoring job 522b3c363499d81ed7922aa30b13e237 from Savepoint 20207 @ 0 for 522b3c363499d81ed7922aa30b13e237 located at abfss://savepoints@storageaccount.dfs.core.windows.net/job-name/savepoint-522b3c-8836a1edc709. 

      Our expectation was that the Operator logs were true, and that it would be restoring from checkpoint. We had to scramble and manually restore from the checkpoint to restore function.



      It's also worth noting I can recreate this issue in a testing environment. The process for doing so is:

      • Boot up HA JobManagers with checkpoints on and savepoint upgrade mode, using adaptive scheduler
      • Make a dummy change to trigger a savepoint.
      • Allow the TaskManagers to process some data and hit the checkpoint interval.
      • Cause the TaskManagers to crash. In our case, we could use up a bunch of memory in the pods and cause it to crash.
      • Observe the Operator logs saying it is restoring from last-state, but watch as the pods instead use the last savepoint.


        Issue Links



              Unassigned Unassigned
              alexdchoffer Alex Hoffer
              0 Vote for this issue
              4 Start watching this issue