Re: [PR] [SPARK-55330][K8S] Make spark.kubernetes.legacy.useReadWriteOnceAccessMode a public config [spark]

["nitrajen (via GitHub)" <gi...@apache.org>]

nitrajen commented on PR #54963:
URL: https://github.com/apache/spark/pull/54963#issuecomment-4183953275

   @dongjoon-hyun Thanks for the review and explanantion!
   
   Understood on the ReadWriteOnce concern - multiple pods on the same node mounting the same PVC.
   
   For users with `claimName: OnDemand` Spark 4.x and get AccessDeniedException creating the blockmgr directory on the PVC mount (same config worked fine on 3.x.). My understanding is with onDemand Spark assigns a unique PVC per executor and explicitly avoids reusing a PVC that's still in use - so sharing is unlikely, though not impossible during the brief window when an old executor is dying and a new one starts. The RWO config is opt-in and default remains ReadWriteOncePod. Apologies if I misunderstood something..
   
   On hostPath - since it's tied to a specific node, if an executor restarts on a different node its data wouldn't be accessible. Would that be a concern for users relying on shuffle recovery across executor restarts?
   
   One alternative is `fsGroupPolicy: File` on the CSI driver, but that's a cluster-wide setting requiring admin access and not always in the hands of the team running Spark jobs.
   
   Happy to close this out for better path or any suggestions..


-- 
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: reviews-unsubscribe@spark.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org