Ngone51 edited a comment on pull request #28656: URL: https://github.com/apache/spark/pull/28656#issuecomment-634776960
@tgravescs The problem is caused by the race condition between `resourcesOffer` and `submitTasks`. If `submitTasks` happens before `resourcesOffer`, especially when there are no executors added to TaskScheduler at all, the TaskSetManager 's `myLocalityLevels` will only have ANY locality level, see: https://github.com/apache/spark/blob/8b4862953a879a9b3ba6f57e669efc383df68b7c/core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala#L218 And then, `resourcesOffer` is called with new executors added to TaskScheduler. Then, `recomputeLocality` will be called because of `executorAdded`. During `recomputeLocality`, the TaskSetManager's `myLocalityLevels` may have PROCESS_LOCAL, NODE_LOCAL, ANY. But the TaskSetManager will stick to the previous locality level, which is ANY, see: https://github.com/apache/spark/blob/8b4862953a879a9b3ba6f57e669efc383df68b7c/core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala#L1107-L1114 As a result, the TaskSetManager can only schedule tasks at ANY level. (I'll update the PR description with more detail explanation later) ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
