[SPARK-25855] Don't use Erasure Coding for event log files

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.4.0
• Fix Version/s: 3.0.0
• Component/s: Spark Core
• Assignee: Imran Rashid
• Reporter: Imran Rashid
• Created: 26/Oct/18 18:07
• Updated: 26/Dec/19 05:08
• Resolved: 31/Oct/18 17:53

Description

While testing spark with hdfs erasure coding (new in hadoop 3), we ran into a bug with the event logs. The main issue was a bug in hdfs (HDFS-14027), but it did make us wonder whether Spark should be using EC for event log files in general. Its a poor choice because EC currently implements hflush() or hsync() as no-ops, which mean you won't see anything in your event logs until the app is complete. That isn't necessarily a bug, but isn't really great. So I think we should ensure EC is always off for event logs.

IIUC there is not a problem with applications which die without properly closing the outputstream. It'll take a while for the NN to realize the client is gone and finish the block, but the data should get there eventually.

Also related are SPARK-24787 & SPARK-19531.

The space savings from EC would be nice as the event logs can get somewhat large, but I think other factors outweigh this.

Issue Links

• is related to: HDFS-14027 DFSStripedOutputStream should implement both hsync methods
• is related to: SPARK-25871 Streaming WAL should not use hdfs erasure coding, regardless of FS defaults
• links to: [Github] Pull Request #22881 (squito)
• links to: [Github] Pull Request #22881 (squito)
• links to: [Github] Pull Request #22916 (squito)
• links to: GitHub Pull Request #22881
• links to: GitHub Pull Request #26998