[SPARK-26418] Only OpenBlocks without any ChunkFetch for one stream will cause memory leak in ExternalShuffleService

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 2.4.0
• Component/s: Shuffle, Spark Core
• Assignee: Unassigned
• Reporter: Wang Shuo
• Created: 20/Dec/18 08:10
• Updated: 17/May/20 18:30
• Resolved: 20/Dec/19 19:32

Description

In current code path,  OneForOneStreamManager holds StreamState in a Map named streams. 

A StreamState is initialized and put into streams when OpenBlocks request received.

One specific StreamState is removed from streams in two scenarios below:

1. The last chunk of a stream is fetched
2. The connection of ChunkFetch is closed

StreamState will never be clean up, if OpenBlocks request is received without and following  ChunkFetch request. This will cause memory leak in server side, which is harmful for long running service such as ExternalShuffleService.

 

Issue Links

• duplicates: SPARK-26604 Register channel for stream request
• links to: [Github] Pull Request #23355 (wangshuo128)
• links to: [Github] Pull Request #23355 (wangshuo128)
• links to: GitHub Pull Request #23355