[YARN-5190] Registering/unregistering container metrics triggered by ContainerEvent and ContainersMonitorEvent are conflict which cause uncaught exception in ContainerMonitorImpl

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Fix Version/s: 2.8.0, 3.0.0-alpha1
• Target Version/s: 2.8.0
• Hadoop Flags: Reviewed
• Assignee: Junping Du
• Reporter: Junping Du
• Created: 01/Jun/16 15:50
• Updated: 25/Oct/19 20:26
• Resolved: 12/Sep/16 17:50

Description

The exception stack is as following:

310735 2016-05-22 01:50:04,554 [Container Monitor] ERROR org.apache.hadoop.yarn.YarnUncaughtExceptionHandler: Thread Thread[Container Monitor,5,main] threw an Exception.
310736 org.apache.hadoop.metrics2.MetricsException: Metrics source ContainerResource_container_1463840817638_14484_01_000010 already exists!
310737         at org.apache.hadoop.metrics2.lib.DefaultMetricsSystem.newSourceName(DefaultMetricsSystem.java:135)
310738         at org.apache.hadoop.metrics2.lib.DefaultMetricsSystem.sourceName(DefaultMetricsSystem.java:112)
310739         at org.apache.hadoop.metrics2.impl.MetricsSystemImpl.register(MetricsSystemImpl.java:229)
310740         at org.apache.hadoop.yarn.server.nodemanager.containermanager.monitor.ContainerMetrics.forContainer(ContainerMetrics.java:212)
310741         at org.apache.hadoop.yarn.server.nodemanager.containermanager.monitor.ContainerMetrics.forContainer(ContainerMetrics.java:198)
310742         at org.apache.hadoop.yarn.server.nodemanager.containermanager.monitor.ContainersMonitorImpl$MonitoringThread.run(ContainersMonitorImpl.java:385)

After YARN-4906, we have multiple places to get ContainerMetrics for a particular container that could cause race condition in registering the same container metrics to DefaultMetricsSystem by different threads. Lacking of proper handling of MetricsException which could get thrown, the exception will could bring down daemon of ContainerMonitorImpl or even whole NM.

Issue Links

• is related to: YARN-5296 NMs going OutOfMemory because ContainerMetrics leak in ContainerMonitorImpl
• is related to: HADOOP-13362 DefaultMetricsSystem leaks the source name when a source unregisters