To reproduce this problem:
- Submit an app which is configured to keep containers across app attempts and should fail after AM finished at first time (am-max-attempts=1).
- App is started with 2 containers running on NM1 node.
- Fail the AM of the application with PREEMPTED exit status which should not count towards max attempt retry but app will fail immediately.
- Used resource of this queue leaks after app fail.
The root cause is the inconsistency of handling app attempt failure between RMAppAttemptImpl$BaseFinalTransition#transition and RMAppImpl$AttemptFailedTransition#transition:
- After app fail, RMAppFailedAttemptEvent will be sent in RMAppAttemptImpl$BaseFinalTransition#transition, if exit status of AM container is PREEMPTED/ABORTED/DISKS_FAILED/KILLED_BY_RESOURCEMANAGER, it will not count towards max attempt retry, so that it will send AppAttemptRemovedSchedulerEvent with keepContainersAcrossAppAttempts=true and RMAppFailedAttemptEvent with transferStateFromPreviousAttempt=true.
- RMAppImpl$AttemptFailedTransition#transition handle RMAppFailedAttemptEvent and will fail the app if its max app attempts is 1.
- CapacityScheduler handles AppAttemptRemovedSchedulerEvent in CapcityScheduler#doneApplicationAttempt, it will skip killing and calling completion process for containers belong to this app, so that queue resource leak happens.