[HADOOP-12321] Make JvmPauseMonitor an AbstractService

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.8.0
• Fix Version/s: 2.9.0, 3.0.0-alpha1
• Target Version/s: 2.9.0
• Assignee: Sunil G
• Reporter: Steve Loughran
• Created: 13/Aug/15 04:58
• Updated: 24/Jan/19 10:43
• Resolved: 06/Dec/15 17:44

Description

The new JVM pause monitor has been written with its own start/stop lifecycle which has already proven brittle to both ordering of operations and, even after HADOOP-12313, is not thread safe (both start and stop are potentially re-entrant).

It also requires every class which supports the monitor to add another field and perform the lifecycle operations in its own lifecycle, which, for all Yarn services, is the YARN app lifecycle (as implemented in Hadoop common)

Making the monitor a subclass of AbstractService and moving the init/start & stop operations in serviceInit(), serviceStart() & serviceStop() methods will fix the concurrency and state model issues, and make it trivial to add as a child to any YARN service which subclasses CompositeService (most the NM and RM apps) will be able to hook up the monitor simply by creating one in the ctor and adding it as a child.

Issue Links

• contains: HDFS-8947 NameNode, DataNode and NFS gateway to support JvmPauseMonitor as a service
• contains: MAPREDUCE-6462 JobHistoryServer to support JvmPauseMonitor as a service
• is depended upon by: HDFS-8947 NameNode, DataNode and NFS gateway to support JvmPauseMonitor as a service
• is depended upon by: MAPREDUCE-6462 JobHistoryServer to support JvmPauseMonitor as a service
• is depended upon by: YARN-4072 ApplicationHistoryServer, WebAppProxyServer, NodeManager and ResourceManager to support JvmPauseMonitor as a service
• is related to: HADOOP-12313 NPE in JvmPauseMonitor when calling stop() before start()
• is related to: HIVE-14150 Hive does not compile against Hadoop-2.9.0-SNAPSHOT
• links to: GitHub Pull Request #54
• links to: GitHub Pull Request #59