Rich-T-kid commented on PR #27252: URL: https://github.com/apache/datafusion/pull/27252#issuecomment-6729488536
## Final comparison: `Dictionary<_, Utf8>` versus `Utf8` Both rounds used `schema_force_view_types=false`; each percentage is RLE off -> on within the same Arrow version. | Suite | Released Arrow | Live Arrow main | Progress | | --- | ---: | ---: | ---: | | TPCH | `+1.74%` | `-1.44%` | `-3.18 pp` | | TPCDS | `-8.39%` | `-10.75%` | `-2.36 pp` | | ClickBench partitioned | `+84.77%` | `+73.58%` | `-11.19 pp` | Live Arrow makes dictionaries faster overall on TPCH and improves the existing TPCDS win. For ClickBench it removes about `13.2%` of the dictionary overhead, but the remaining `+73.58%` regression is still large. The ClickBench progress is concentrated in Q34, Q33, Q12, Q13, Q5, and Q14. Q22 remains about `+150%`, and Q23 moves from about `+142%` to `+146%`, so Q22/Q23 remain the clearest bottlenecks. For context, disabling views on released Arrow changes the earlier RLE deltas from `+8.69% / +0.38% / +109.52%` to `+1.74% / -8.39% / +84.77%` for TPCH/TPCDS/ClickBench. That confirms `Utf8View` is a materially stronger baseline than plain `Utf8`. -- 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]
