[YARN-6714] IllegalStateException while handling APP_ATTEMPT_REMOVED event when async-scheduling enabled in CapacityScheduler

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.9.0, 3.0.0-alpha3
• Fix Version/s: 2.9.0, 3.0.0-beta1
• Hadoop Flags: Reviewed
• Assignee: Tao Yang
• Reporter: Tao Yang
• Created: 15/Jun/17 09:56
• Updated: 17/Jul/17 07:58
• Resolved: 17/Jul/17 07:58

Description

Currently in async-scheduling mode of CapacityScheduler, after AM failover and unreserve all reserved containers, it still have chance to get and commit the outdated reserve proposal of the failed app attempt. This problem happened on an app in our cluster, when this app stopped, it unreserved all reserved containers and compared these appAttemptId with current appAttemptId, if not match it will throw IllegalStateException and make RM crashed.

Error log:

2017-06-08 11:02:24,339 FATAL [ResourceManager Event Processor] org.apache.hadoop.yarn.server.resourcemanager.ResourceManager: Error in handling event type APP_ATTEMPT_REMOVED to the scheduler
java.lang.IllegalStateException: Trying to unreserve  for application appattempt_1495188831758_0121_000002 when currently reserved  for application application_1495188831758_0121 on node host: node1:45454 #containers=2 available=... used=...
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode.unreserveResource(FiCaSchedulerNode.java:123)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp.unreserve(FiCaSchedulerApp.java:845)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.LeafQueue.completedContainer(LeafQueue.java:1787)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.completedContainerInternal(CapacityScheduler.java:1957)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.AbstractYarnScheduler.completedContainer(AbstractYarnScheduler.java:586)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.doneApplicationAttempt(CapacityScheduler.java:966)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.handle(CapacityScheduler.java:1740)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.handle(CapacityScheduler.java:152)
        at org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$SchedulerEventDispatcher$EventProcessor.run(ResourceManager.java:822)
        at java.lang.Thread.run(Thread.java:834)

When async-scheduling enabled, CapacityScheduler#doneApplicationAttempt and CapacityScheduler#tryCommit both need to get write_lock before executing, so we can check the app attempt state in commit process to avoid committing outdated proposals.