Rich-T-kid commented on PR #26789: URL: https://github.com/apache/datafusion/pull/26789#issuecomment-6849396171
20-iteration follow-up: - [Disabled → enabled](https://github.com/apache/datafusion/pull/26789#issuecomment-6848454261): primary table is 23 faster / 0 slower; mean table is 26 faster / 0 slower. Mean suite time drops from 9256.19 ms to 8882.08 ms (-4.04%). - [Enabled → disabled](https://github.com/apache/datafusion/pull/26789#issuecomment-6848455285): Q47 and Q57 still favor enabled even from the first position. Q47 is 208.77 vs 216.09 ms; Q57 is 155.44 vs 166.45 ms. - [Enabled → enabled A/A](https://github.com/apache/datafusion/pull/26789#issuecomment-6848800764): identical code and flags report 12 mean improvements and one 1.07x regression, with the second position 2.67% faster overall. This confirms a substantial runner-order effect. Balancing the two flag orientations gives: - Q47: 11.3% lower mean time with the optimizer enabled. - Q57: 7.7% lower mean time with the optimizer enabled. - Whole suite: 1.18% lower mean time with the optimizer enabled. The reverse-only flags (Q40, Q68, Q74, Q76, Q78) do not rewrite and have byte-identical physical plans. Q74 was the only residual concern after correcting for position, so I isolated it at exact head with Q1-Q73 preconditioning and 160 alternating pairs. Enabled changed physical planning by +0.006 ms (+0.13%), execution by -0.030 ms (-0.03%), and total time by -0.030 ms (-0.03%). The hosted Q74 gap does not reproduce. Conclusion: Q47/Q57 are repeatable improvements, and there is no reproducible regression attributable to the optimizer. The PR head is restored to `678d97de914fd0e55fbc1f2d5eea6ac67a916dff`; the benchmark-only wrapper commits are not in the PR history. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
