alamb commented on PR #10852: URL: https://github.com/apache/arrow-rs/pull/10852#issuecomment-5443608575
I ran an end-to-end DataFusion ClickBench Q25 validation of this PR (via the 59_maintenance backport in #10885), as part of the filter-pushdown-by-default effort in apache/datafusion#24426. Setup: `--profile=profiling` builds of `datafusion-cli` against `hits_partitioned`, filter pushdown enabled; 3 interleaved rounds × 10 measured runs per binary, stdev ≤3.5ms. Q25 (`SELECT "SearchPhrase" FROM hits WHERE "SearchPhrase" <> '' ORDER BY "SearchPhrase" LIMIT 10`) is the query whose profile motivated #10776: a ~15%-selective scattered selection over a large string column. | | avg ms | vs pushdown-off baseline | |---|---|---| | pushdown on, parquet 59.2.0 | 113.1 | 1.34x | | pushdown on, + this PR | 101.9 | 1.21x | | pushdown off (baseline) | 84.2 | 1.00x | **-11.2ms, about 40% of the total pushdown penalty on this query** — right in line with the ~100ms of per-query selection-construction CPU the samply profiles attributed to the booleans → selectors → mask round trip (spread over ~10 effective cores). Details and the full five-configuration matrix are in https://github.com/apache/datafusion/pull/24426#issuecomment-5443606081. The remaining gap on this query shape is dominated by the `CachedArrayReader` replay concat (#10774). Thanks @hhhizzz — this is a real, measurable end-to-end win. -- 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]
