cloud-fan commented on pull request #35460: URL: https://github.com/apache/spark/pull/35460#issuecomment-1047487230
This is really a hard problem and rerunning the entire stage is more of a compromise. In a large enough cluster, we may always see task failures when running a stage, and rerunning the entire stage may never succeed. That's why in Spark SQL, we don't really rely on the `DeterministicLevel` framework, but by default we sort before repartition to fix the correctness issue. I think we should either have reliable shuffle storage (AFAIK there are several third-party remote shuffle services) so that fetch failure never happens, or we reject such queries. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
