[HIVE-20393] Semijoin Reduction : markSemiJoinForDPP behaves inconsistently

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Assignee: Deepak Jaiswal
• Reporter: Deepak Jaiswal
• Created: 15/Aug/18 07:00
• Updated: 16/Aug/18 07:59
• Resolved: 16/Aug/18 07:59

Description

markSemiJoinForDPP has multiple issues,

 

• Uses map tsOps which is wrong as it disallows going thru same TS which may have filters from more than 1 semijoin edges. This results in inconsistent plans for same query as semijoin edges may be processed in different order each time.
• Uses getColumnExpr() which is not as robust as extractColumn() thus resulting in NPEs.
• The logic to mark an edge useful when NPE is hit may end up having bad edge.

cc gopalv

Issue Links

• links to: RB