[HIVE-24670] DeleteReaderValue should not allocate empty vectors for delete delta files

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.0.0-alpha-1
• Labels: pull-request-available
• Assignee: Ádám Szita
• Reporter: Ádám Szita
• Created: 20/Jan/21 14:38
• Updated: 17/Nov/22 08:48
• Resolved: 26/Jan/21 12:54

Description

If delete delta caching is turned off, the plain record reader inside DeleteReaderValue allocates a batch with a schema that is equivalent to that of an insert delta.

This is unnecessary as the struct part in a delete delta file is always empty. In cases where we have many delete delta files (e.g. due to compaction failures) and a wide table definition (e.g. 200+ cols) this puts a significant amount of memory pressure on the executor, while these empty structures will never be filled or otherwise utilized.

I propose we specify an ACID schema with an empty struct part to this record reader to counter this.

Issue Links

• links to: GitHub Pull Request #1894