[SPARK-28743] YarnShuffleService leads to NodeManager OOM because ChannelOutboundBuffer has too many entries

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.3.0
• Component/s: Shuffle, Spark Core
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Jiandan Yang
• Created: 15/Aug/19 08:55
• Updated: 25/May/21 01:53
• Resolved: 25/May/21 01:40

Description

NodeManager heap size is 4G, io.netty.channel.ChannelOutboundBuffer$Entry occupied about 2.8G by looking at Histogram of Mat, and those Entries were hold by ChannelOutboundBuffer by looking at dominator_tree of mat. By analyzing  one fo ChannelOutboundBuffer object, I found there were 248867 entries in the object of ChannelOutboundBuffer (ChannelOutboundBuffer#flushed=248867), and  ChannelOutboundBuffer#totalPengdingSize=23891232 which is more than highwaterMark(64K), and unwritable=1 meaning sending buffer was full.  But ChannelHandler seems not check unwritable flag when write message, and finally NodeManager occurs OOM.

Histogram:

dominator_tree: