[YARN-1795] After YARN-713, using FairScheduler can cause an InvalidToken Exception for NMTokens

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Duplicate
• Affects Version/s: 2.4.0
• Target Version/s: 2.4.0
• Assignee: Robert Kanter
• Reporter: Robert Kanter
• Created: 07/Mar/14 01:28
• Updated: 17/Mar/14 20:40
• Resolved: 17/Mar/14 20:39

Description

Running the Oozie unit tests against a Hadoop build with YARN-713 causes many of the tests to be flakey. Doing some digging, I found that they were failing because some of the MR jobs were failing; I found this in the syslog of the failed jobs:

2014-03-05 16:18:23,452 INFO [AsyncDispatcher event handler] org.apache.hadoop.mapreduce.v2.app.job.impl.TaskAttemptImpl: Diagnostics report from attempt_1394064846476_0013_m_000000_0: Container launch failed for container_1394064846476_0013_01_000003 : org.apache.hadoop.security.token.SecretManager$InvalidToken: No NMToken sent for 192.168.1.77:50759
       at org.apache.hadoop.yarn.client.api.impl.ContainerManagementProtocolProxy$ContainerManagementProtocolProxyData.newProxy(ContainerManagementProtocolProxy.java:206)
       at org.apache.hadoop.yarn.client.api.impl.ContainerManagementProtocolProxy$ContainerManagementProtocolProxyData.<init>(ContainerManagementProtocolProxy.java:196)
       at org.apache.hadoop.yarn.client.api.impl.ContainerManagementProtocolProxy.getProxy(ContainerManagementProtocolProxy.java:117)
       at org.apache.hadoop.mapreduce.v2.app.launcher.ContainerLauncherImpl.getCMProxy(ContainerLauncherImpl.java:403)
       at org.apache.hadoop.mapreduce.v2.app.launcher.ContainerLauncherImpl$Container.launch(ContainerLauncherImpl.java:138)
       at org.apache.hadoop.mapreduce.v2.app.launcher.ContainerLauncherImpl$EventProcessor.run(ContainerLauncherImpl.java:369)
       at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
       at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
       at java.lang.Thread.run(Thread.java:744)

I did some debugging and found that the NMTokenCache has a different port number than what's being looked up. For example, the NMTokenCache had one token with address 192.168.1.77:58217 but ContainerManagementProtocolProxy.java:119 is looking for 192.168.1.77:58213. The 58213 address comes from ContainerLauncherImpl's constructor. So when the Container is being launched it somehow has a different port than when the token was created.

Any ideas why the port numbers wouldn't match?

Update: This also happens in an actual cluster, not just Oozie's unit tests

Issue Links

• duplicates: YARN-1839 Capacity scheduler preempts an AM out. AM attempt 2 fails to launch task container with SecretManager$InvalidToken: No NMToken sent
• is broken by: YARN-713 ResourceManager can exit unexpectedly if DNS is unavailable