[SPARK-26712] Single disk broken causing YarnShuffleSerivce not available

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Won't Fix
• Affects Version/s: 2.1.0, 2.4.0
• Component/s: Shuffle, Spark Core
• Assignee: Unassigned
• Reporter: liupengcheng
• Created: 24/Jan/19 03:56
• Updated: 17/May/20 18:30
• Resolved: 26/Feb/19 15:40

Description

Currently, `ExecutorShuffleInfo` can be recovered from file if NM recovery enabled, however, the recovery file is under a single directory, which may be unavailable if disk broken. So if a NM restart happen(may be caused by kill or some reason), the shuffle service can not start even if there are executors on the node.

This may finally cause job failures(if node or executors on it not blacklisted), or at least, it will cause resource waste.(shuffle from this node always failed.)

For long running spark applications, this problem may be more serious.

So I think we should support multi directories(multi disk) for this recovery. and change to good directory when the disk of current directory is broken.

Issue Links

• links to: GitHub Pull Request #23647