[
https://issues.apache.org/jira/browse/YARN-2416?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jason Lowe updated YARN-2416:
-----------------------------
Priority: Critical (was: Major)
Bumping the priority of this since easily leads to infinite job hangs:
# AM races ahead of launch event and tries to register.
# ApplicationMasterService ends up completing the registration from its point
of view and fires event to app state machine
# App state machine still thinks it hasn't launched yet, so it throws the
stacktrace listed above and remains in the ALLOCATED state.
# Eventually launch even arrives and we move to the LAUNCHED state.
# We can never move to the RUNNING state since the app and
ApplicationMasterService both think we've registered but the app state machine
does not.
# App tries to unregister. Since ApplicationMasterService thinks it has
registered, it tickles the expiration watchdog and sends the event to the app
state machine. App is told unregister was false because it hasn't been saved
in a terminal state
# App state machine barfs and ignores unregister event since app is not RUNNING
(and never will be)
# Rinse, repeat unregistration forever until app is manually killed. The RM
will never expire the app because ApplicationMasterService keeps tickling the
watchdog on each unregister attempt.
bq. Now AM.register and Event.Launched is asynchronous. If need to change it
to be synchronous?
This is not easy to do. There could be a network cut between the RM and the NM
or a large GC event just as the NM tries to send the RPC response to the RM,
but meanwhile the AM proceeds to launch and ends up registering to the RM just
as the NM RPC response arrives. The only way to enforce ordering there is to
temporarily reject AM registrations until the state machine has seen the launch
event. I'd much prefer to just fix the state machine to expect these two
events to race inherently as originally suggested.
> InvalidStateTransitonException in ResourceManager if AMLauncher does not
> receive response for startContainers() call in time
> ----------------------------------------------------------------------------------------------------------------------------
>
> Key: YARN-2416
> URL: https://issues.apache.org/jira/browse/YARN-2416
> Project: Hadoop YARN
> Issue Type: Bug
> Components: resourcemanager
> Affects Versions: 2.4.0
> Reporter: Jian Fang
> Assignee: Jonathan Eagles
> Priority: Critical
>
> AMLauncher calls startContainers(allRequests) to launch a container for
> application master. Normally, the call comes back immediately so that the
> RMAppAttempt changes its state from ALLOCATED to LAUNCHED.
> However, we do observed that in some cases, the RPC call came back very late
> but the AM container was already started. Because the RMAppAttempt stuck in
> ALLOCATED state, once resource manager received the REGISTERED event from the
> application master, it threw InvalidStateTransitonException as follows.
> 2014-07-05 08:59:05,021 ERROR
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl:
> Can't handle this event at current state
> org.apache.hadoop.yarn.state.InvalidStateTransitonException: Invalid event:
> REGISTERED at ALLOCATED
> at
> org.apache.hadoop.yarn.state.StateMachineFactory.doTransition(StateMachineFactory.java:305)
> at
> org.apache.hadoop.yarn.state.StateMachineFactory.access$300(StateMachineFactory.java:46)
> at
> org.apache.hadoop.yarn.state.StateMachineFactory$InternalStateMachine.doTransition(StateMachineFactory.java:448)
> at
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl.handle(RMAppAttemptImpl.java:652)
> at
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl.handle(RMAppAttemptImpl.java:106)
> at
> org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$ApplicationAttemptEventDispatcher.handle(ResourceManager.java:752)
> at
> org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$ApplicationAttemptEventDispatcher.handle(ResourceManager.java:733)
> at
> org.apache.hadoop.yarn.event.AsyncDispatcher.dispatch(AsyncDispatcher.java:173)
> at
> org.apache.hadoop.yarn.event.AsyncDispatcher$1.run(AsyncDispatcher.java:106)
> at java.lang.Thread.run(Thread.java:744)
> For subsequent STATUS_UPDATE and CONTAINER_ALLOCATED events for this job,
> resource manager kept throwing InvalidStateTransitonException.
> 2014-07-05 08:59:06,152 ERROR
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl:
> Can't handle this event at current state
> org.apache.hadoop.yarn.state.InvalidStateTransitonException: Invalid event:
> STATUS_UPDATE at ALLOCATED
> at
> org.apache.hadoop.yarn.state.StateMachineFactory.doTransition(StateMachineFactory.java:305)
> at
> org.apache.hadoop.yarn.state.StateMachineFactory.access$300(StateMachineFactory.java:46)
> at
> org.apache.hadoop.yarn.state.StateMachineFactory$InternalStateMachine.doTransition(StateMachineFactory.java:448)
> at
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl.handle(RMAppAttemptImpl.java:652)
> at
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl.handle(RMAppAttemptImpl.java:106)
> at
> org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$ApplicationAttemptEventDispatcher.handle(ResourceManager.java:752)
> at
> org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$ApplicationAttemptEventDispatcher.handle(ResourceManager.java:733)
> at
> org.apache.hadoop.yarn.event.AsyncDispatcher.dispatch(AsyncDispatcher.java:173)
> at
> org.apache.hadoop.yarn.event.AsyncDispatcher$1.run(AsyncDispatcher.java:106)
> at java.lang.Thread.run(Thread.java:744)
> 2014-07-05 08:59:07,779 INFO
> org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainerImpl:
> container_1404549222428_0001_02_000002 Container Transitioned from NEW to
> ALLOCATED
> 2014-07-05 08:59:07,779 ERROR
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl:
> Can't handle this event at current state
> org.apache.hadoop.yarn.state.InvalidStateTransitonException: Invalid event:
> CONTAINER_ALLOCATED at ALLOCATED
> at
> org.apache.hadoop.yarn.state.StateMachineFactory.doTransition(StateMachineFactory.java:305)
> at
> org.apache.hadoop.yarn.state.StateMachineFactory.access$300(StateMachineFactory.java:46)
> at
> org.apache.hadoop.yarn.state.StateMachineFactory$InternalStateMachine.doTransition(StateMachineFactory.java:448)
> at
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl.handle(RMAppAttemptImpl.java:652)
> at
> org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl.handle(RMAppAttemptImpl.java:106)
> at
> org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$ApplicationAttemptEventDispatcher.handle(ResourceManager.java:752)
> at
> org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$ApplicationAttemptEventDispatcher.handle(ResourceManager.java:733)
> at
> org.apache.hadoop.yarn.event.AsyncDispatcher.dispatch(AsyncDispatcher.java:173)
> at
> org.apache.hadoop.yarn.event.AsyncDispatcher$1.run(AsyncDispatcher.java:106)
> at java.lang.Thread.run(Thread.java:744)
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
