Hi hackers! I'd stumbled upon the discussion [1] on TPC-DS query performance, looked into code that caused this behavior and saw that partial hash, merge and nested loop paths are discarded regardless of costs if they have more pathkeys.
I've excluded the pathkey chain length condition from the precheck function and added passing precalculated startup cost in addition to total cost, and it seems to produce a more effective plan for case in [1]. Please check the attached patch. I'm very interested if my assumption is correct or not. [1] https://www.postgresql.org/message-id/flat/SEZPR06MB649422CDEBEBBA3915154EE58A232%40SEZPR06MB6494.apcprd06.prod.outlook.com -- Regards, Nikita Malakhov Postgres Professional The Russian Postgres Company https://postgrespro.ru/
v1-0001-ppath-precheck.patch
Description: Binary data
