alexanderbianchi opened a new pull request, #130: URL: https://github.com/apache/datafusion-iceberg/pull/130
## What does this measure? At 400 distinct string keys, the Iceberg reader change reduced median end-to-end DataFusion query CPU from **149.09 ms to 14.40 ms (90.3%)** and wall time from **148.97 ms to 14.27 ms (10.44× faster)**. This draft benchmarks the hash-set change in [apache/iceberg-rust#3739](https://github.com/apache/iceberg-rust/pull/3739). The query selects `dd_resource_key` and filters that same column with a literal `IN` list. Both builds read the same 175,000-row snapshot across seven NYC taxi Parquet files. Each trip has one distinct, synthetic 63–68 byte string key. A 400-key query matches 400 trips spread across all seven files. ## Results Medians of ten warm, release-build queries per cell on an Apple M5 Pro. CPU is process-wide. Both measurements include SQL parsing, planning, execution and result collection. | Distinct keys | Baseline CPU (ms) | Patched CPU (ms) | Baseline wall (ms) | Patched wall (ms) | Wall speedup | | ---: | ---: | ---: | ---: | ---: | ---: | | 1 | 5.32 | 5.49 | 5.22 | 5.36 | 0.97× | | 16 | 11.20 | 10.93 | 11.06 | 10.81 | 1.02× | | 31 | 16.62 | 16.31 | 16.49 | 16.19 | 1.02× | | 32 | 17.05 | 10.76 | 16.89 | 10.64 | 1.59× | | 100 | 40.90 | 13.73 | 40.76 | 13.58 | 3.00× | | 369 | 139.83 | 13.00 | 139.67 | 12.87 | 10.85× | | 400 | 149.09 | 14.40 | 148.97 | 14.27 | 10.44× | | 1,000 | 375.52 | 13.77 | 375.42 | 13.62 | 27.55× | The hash path starts at 32 literals. At 400 keys, median planning CPU was 1.48 ms before and 1.31 ms after. Median execution CPU fell from 147.61 ms to 13.10 ms. Both executed plans pushed the same 400 keys into Iceberg, emitted 400 rows from the scan, and passed all 400 through DataFusion’s residual `FilterExec`. [Baseline plan](https://github.com/alexanderbianchi/datafusion-iceberg-upstream/blob/6bbb474eec0f5c3a5aa6768cd0a3aadfb2a065d7/benchmark-results/before-400-plan.txt) · [Patched plan](https://github.com/alexanderbianchi/datafusion-iceberg-upstream/blob/6bbb474eec0f5c3a5aa6768cd0a3aadfb2a065d7/benchmark-results/after-400-plan.txt) · [Raw samples](https://github.com/alexanderbianchi/datafusion-iceberg-upstream/blob/6bbb474eec0f5c3a5aa6768cd0a3aadfb2a065d7/benchmark-results/samples.csv) ## Reproduction and validation - Use DataFusion 55.2.0 and Arrow 59.3.0. Both builds use the same Arrow 59-compatible reader baseline. For the patched build, cherry-pick the reader change onto that baseline. [Baseline commit](https://github.com/DataDog/iceberg-rust/commit/9354b42ad8c28b9f3df2ae4c3e4931beab0be12b) · [Reader change](https://github.com/apache/iceberg-rust/commit/a16ffe709c258595895f5d3c9f00d34b3abedc9f) - Materialize the seven Parquet files with Git LFS from the January 8–14 taxi fixture. From this PR checkout, run `python3 dev/bench_taxi_in_list.py <taxi-data-directory> <patched-iceberg-checkout> target/taxi-results`. [Pinned fixture](https://github.com/apache/datafusion-iceberg/tree/82cea1462bff959021b881311c9ef9f4f22328af/crates/datafusion/testdata/iceberg/taxi) - The runner used one local query at a time and alternated builds in baseline/patched/patched/baseline order. All 160 timed queries returned exactly the expected keys. The baseline passed workspace tests and Clippy, and the patched build passed all 41 connector integration tests. [Full methodology](https://github.com/alexanderbianchi/datafusion-iceberg-upstream/blob/6bbb474eec0f5c3a5aa6768cd0a3aadfb2a065d7/benchmark-results/README.md) -- 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]
