[SPARK-30271] dynamic allocation won't release some executor in some case.

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.4.0
• Component/s: Scheduler, Spark Core
• Assignee: Unassigned
• Reporter: angerszhu
• Created: 16/Dec/19 14:14
• Updated: 09/Oct/20 07:18
• Resolved: 09/Oct/20 07:18

Description

Case :
max executor 5
min executor 0
idle time 5s

stage-1 10 tasks run in 5 executors.
If stage-1 finished in 5 all executors, all executor added to `removeTimes` when taskEnd event.
After 5s, start release process, since stage-2 have 20 tasks, then executor won't be removed since existing executor num < executorTargetNum., and executor will be removed from `removeTimes`.
But if task won't be scheduled to all these executors, if executor-1 won't have task to run in it, it won't be put into `removeTimes` and if there are no more tasks, executor won't be removed forever