romina-nikolova opened a new issue, #72803:
URL: https://github.com/apache/airflow/issues/72803

   ### Under which category would you file this issue?
   
   Providers
   
   ### Apache Airflow version
   
   3.3.1
   
   ### What happened and how to reproduce it?
   
   KubernetesExecutor issues 2-3 delete_pod API calls for the same finished 
worker pod instead of one. The first call succeeds; the rest return 404 from 
the Kubernetes API. All calls come from the same scheduler process for the same 
TaskInstanceKey, milliseconds apart.
   
   Observed on a production-scale cluster: over a 15-minute sample, 306 pods 
received 3 delete calls, 15 received 2, and only 4 received the intended 1. 
That's ~1.47 delete calls per pod and a 67% 404 rate, generating a large amount 
of unnecessary kube-apiserver audit-log volume (and associated CloudWatch cost 
on EKS) at scale. Mapped tasks (non-null map_index) appear to be the common 
factor in every case we captured.
   
   Root cause
   
   In the result-processing path (kubernetes_executor.py / 
kubernetes_executor_utils.py in apache-airflow-providers-cncf-kubernetes), 
delete_pod is invoked before the executor's self.running dedup check runs for 
that TaskInstanceKey. Because the same task-completion event gets processed 
more than once before the dedup check has a chance to short-circuit it, the pod 
delete is issued redundantly on each pass.
   
   ### What you think should happen instead?
   
   The dedup check against self.running should happen (or the running-set be 
updated) before delete_pod is called, so a given TaskInstanceKey only ever 
triggers one delete call, regardless of how many times its completion event is 
processed.
   
   ### Operating System
   
   _No response_
   
   ### Deployment
   
   None
   
   ### Apache Airflow Provider(s)
   
   _No response_
   
   ### Versions of Apache Airflow Providers
   
   apache-airflow-providers-cncf-kubernetes==10.21.0
   
   ### Official Helm Chart version
   
   Not Applicable
   
   ### Kubernetes Version
   
   _No response_
   
   ### Helm Chart configuration
   
   _No response_
   
   ### Docker Image customizations
   
   _No response_
   
   ### Anything else?
   
   _No response_
   
   ### Are you willing to submit PR?
   
   - [ ] Yes I am willing to submit a PR!
   
   ### Code of Conduct
   
   - [x] I agree to follow this project's [Code of 
Conduct](https://github.com/apache/airflow/blob/main/CODE_OF_CONDUCT.md)
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to