Github user jerryshao commented on the issue: https://github.com/apache/spark/pull/11205 @vanzin , in the current code `stageIdToTaskIndices` cannot be used to track number of running tasks, because this structure doesn't remove task index from itself when task is finished successfully. Yes `isExecutorIdle` is used to take care of executor idle, but the way to identify whether executor is idle is not robust enough. In this scenario, when stage is aborted because of max task failures, some task end event will be missing, so using number of tasks per executor will lead to residual data, and makes executor always be busy.
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org