geoffreyclaude commented on issue #22849: URL: https://github.com/apache/datafusion/issues/22849#issuecomment-4660457988
> I will open a PR shortly. [@geoffreyclaude](https://github.com/geoffreyclaude) since you were the original author of [#15563](https://github.com/apache/datafusion/pull/15563) you might want to have a look Thanks for the heads-up. I ran the `sort-tpch --sorted --limit 10` benchmark from #15563 before/after the suspected regression point. Setup: - Before: `0c3bb78e24722e2a4f19cdc9a76f1c498949f5be` - After: `1429c92474238a91a09f1cd4a68c19d03329b6a7` (#15770 merge commit) - Command: ```bash dfbench sort-tpch --sorted --limit 10 --iterations 5 \ --path /tmp/df-topk-bench-data/tpch_sf1 \ -o <result>.json ``` Mean timings: | Query | Before | After | Change | |---|---:|---:|---:| | Q1 | 56.63 ms | 26.51 ms | 2.14x faster | | Q2 | 14.98 ms | 8.83 ms | 1.70x faster | | Q3 | 82.64 ms | 64.13 ms | 1.29x faster | | Q4 | 5.00 ms | 16.44 ms | 3.29x slower | | Q5 | 18.34 ms | 12.55 ms | 1.46x faster | | Q6 | 30.69 ms | 24.49 ms | 1.25x faster | | Q7 | 69.32 ms | 62.09 ms | 1.12x faster | | Q8 | 5.60 ms | 27.42 ms | 4.90x slower | | Q9 | 7.77 ms | 67.61 ms | 8.71x slower | | Q10 | 10.20 ms | 60.39 ms | 5.92x slower | | Q11 | 57.23 ms | 48.43 ms | 1.18x faster | Total mean time went from `358.40 ms` to `418.88 ms` (`1.17x` slower overall). The important part is the shape: Q4, Q8, Q9, and Q10 are repeatedly much slower after #15770, while several other queries get faster. That matches this issue pretty well: the regression seems specific to the partial-prefix TopK early-exit path, not a general TopK slowdown. Looking forward to your fix PR! -- 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]
