[MAPREDUCE-4733] Reducer can fail to make progress during shuffle if too many reducers complete consecutively

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.23.3
• Fix Version/s: 2.0.3-alpha, 0.23.5
• Component/s: applicationmaster, mrv2
• Target Version/s: 2.0.3-alpha, 0.23.5
• Hadoop Flags: Reviewed
• Assignee: Jason Darrell Lowe
• Reporter: Jason Darrell Lowe
• Created: 18/Oct/12 22:30
• Updated: 02/May/13 02:29
• Resolved: 19/Oct/12 20:22

Description

TaskAttemptListenerImpl implements getMapCompletionEvents by calling Job.getTaskAttemptCompletionEvents with the same fromEvent and maxEvents passed in from the reducer and then filtering the result for just map events. We can't filter the task completion event list and expect the caller's "window" into the list to match up. As soon as a reducer event appears in the list it means we are redundantly sending map completion events that were already seen by the reducer.

Worst case the reducer will hang if all of the events in the requested window are reducer events. In that case zero events will be reported back to the caller and it won't bump up fromEvent on the next call. Reducer then never sees the final map completion events needed to complete the shuffle. This could happen in a case where all maps complete, more than MAX_EVENTS reducers complete consecutively, but some straggling reducers get fetch failures and cause a map to be restarted.

Issue Links

• is depended upon by: MAPREDUCE-4730 AM crashes due to OOM while serving up map task completion events