[YARN-5937] stop-yarn.sh is not able to gracefully stop node managers

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.0.0-alpha2
• Labels: script
• Target Version/s: 3.0.0-alpha2
• Hadoop Flags: Reviewed
• Assignee: Weiwei Yang
• Reporter: Weiwei Yang
• Created: 28/Nov/16 16:46
• Updated: 10/Jan/17 10:33
• Resolved: 10/Jan/17 10:33

Description

stop-yarn.sh always gives following output

./sbin/stop-yarn.sh
Stopping resourcemanager
Stopping nodemanagers
<NM_HOST>: WARNING: nodemanager did not stop gracefully after 5 seconds: Trying to kill with kill -9
<NM_HOST>: ERROR: Unable to kill 18097

this was because resource manager is stopped before node managers, when the shutdown hook manager tries to gracefully stop NM services, NM needs to unregister with RM, and it gets timeout as NM could not connect to RM (already stopped). See log (stop RM then run kill <nm_pid>)

16/11/28 08:26:43 ERROR nodemanager.NodeManager: RECEIVED SIGNAL 15: SIGTERM
...
16/11/28 08:26:53 WARN util.ShutdownHookManager: ShutdownHook 'CompositeServiceShutdownHook' timeout, java.util.concurrent.TimeoutException
java.util.concurrent.TimeoutException
	at java.util.concurrent.FutureTask.get(FutureTask.java:205)
	at org.apache.hadoop.util.ShutdownHookManager$1.run(ShutdownHookManager.java:67)
...
	at org.apache.hadoop.yarn.server.nodemanager.NodeStatusUpdaterImpl.unRegisterNM(NodeStatusUpdaterImpl.java:291)
...
16/11/28 08:27:13 ERROR util.ShutdownHookManager: ShutdownHookManger shutdown forcefully.

the shutdown hooker has a default of 10s timeout, so if RM is stopped before NMs, they always took more than 10s to stop (in java code). However stop-yarn.sh only gives 5s timeout, so NM is always killed instead of stopped.

It would make sense to stop NMs before RMs in this script, in a graceful way.