[HADOOP-3415] JobEndNotifier isnt synchronized, doesnt check state before acting

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Invalid
• Affects Version/s: 0.19.0
• Assignee: Unassigned
• Reporter: Steve Loughran
• Created: 19/May/08 13:11
• Updated: 29/Apr/10 08:06
• Resolved: 10/Jul/08 23:12

Description

JobEndNotifier is pretty hazardous inside.

1. the static startNotifier isnt synchronized, and doesnt check for being already running before it creates a new worker thread. It should be sycnhronized and a no-op if there is a live thread.

2. stopNotifier() should be a no-op if already stopped. It MUST NOT call thread.interrupt() in such a state, as thread may be null.

3. the registerNotification method also assumes that the static queue is non null.

Things would be a lot safer by making this class part of a JobTracker, not a singleton with static methods, as then you could more safely make assumptions about object state. This would not only eliminate a lot of reentrancy problems, but tie the life of the notifier to that of its owner, the JobTracker.

Issue Links

• is part of: HDFS-326 Add a lifecycle interface for Hadoop components: namenodes, job clients, etc.