[SPARK-20715] MapStatuses shouldn't be redundantly stored in both ShuffleMapStage and MapOutputTracker

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.3.0
• Fix Version/s: 2.2.3, 2.3.0
• Component/s: Scheduler, Shuffle, Spark Core
• Target Version/s: 2.3.0
• Assignee: Josh Rosen
• Reporter: Josh Rosen
• Created: 11/May/17 22:33
• Updated: 17/May/20 17:46
• Resolved: 12/Jun/17 15:19

Description

Today the MapOutputTracker and ShuffleMapStage both maintain their own copies of MapStatuses. This creates the potential for bugs in case these two pieces of state become out of sync.

I believe that we can improve our ability to reason about the code by storing this information only in the MapOutputTracker. This can also help to reduce driver memory consumption.

I will provide more details in my PR, where I'll walk through the detailed arguments as to why we can take these two different metadata tracking formats and consolidate without loss of performance or correctness.

Issue Links

• is related to: SPARK-21444 Fetch failure due to node reboot causes job failure
• links to: [Github] Pull Request #17955 (JoshRosen)
• links to: [Github] Pull Request #22382 (bersprockets)
• links to: [Github] Pull Request #22382 (bersprockets)