sunchao commented on PR #25584: URL: https://github.com/apache/datafusion/pull/25584#issuecomment-6758068322
Rebased onto `c8265b6872` and addressed the review feedback in `9608e5bd01`. - The join recognizes the comparison once per stream. After the first two probes, it checks doubling prefixes, including for singleton outer slices. It keeps the original inner group if a partial scan stops early. - A completed summary owns compact Arrow scalars and resizes the group reservation to their combined size. A refused resize preserves the original buffer. The cached charge is freed at key completion. An unchanged long string stays in place when later batches lose the min/max comparison. - The new tests exercise singleton reuse, partial scans followed by a witness in a later batch, reservation refusal and release, an upstream error with a live summary, and compound predicates whose conditions need a common witness. Fork CI passed all 27 executed jobs. [CI run](https://github.com/sunchao/arrow-datafusion/actions/runs/42738341311) - On identical fixtures, the Utf8 third-row case took 0.632 ms on base and 0.627 ms on head. The latest-version SQL query took 38.00 ms and 6.50 ms. The original TPC runs used hash joins; Q27 exercises this sort-merge change directly. The PR description includes the slower unfiltered result from the first run, the eight-pair follow-up, and Q13 output parity. [Matched benchmarks](https://github.com/sunchao/arrow-datafusion/actions/runs/42738952097) ยท [Follow-up controls](https://github.com/sunchao/arrow-datafusion/actions/runs/42743163590) -- 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]
