sunchao commented on PR #11279: URL: https://github.com/apache/arrow-rs/pull/11279#issuecomment-6004482826
@andygrove Thanks for verifying the large-value case. I addressed the follow-ups; the current head is `c8ee97880e46af9257a41e8df355bd9bcf23f238`: 1. Byte-view benchmarks now have a separate Criterion configuration: 20 samples, 100 ms warmup, and 300 ms measurement. All 42 workloads remain, and the existing non-byte-view benchmarks keep their defaults. Command-line timing options still override the short defaults. A full run of all 168 cases with no timing overrides completed in 91.41 seconds on the measurement host; a separate run verified 300 ms warmup and 700 ms measurement CLI overrides. The benchmark-only companion #11346 has the same configuration. 2. Added `Interleaver::interleave_record_batch`. It applies the options column by column, preserves the first schema and its metadata, requires equal input schemas, and handles zero-column batches with the selected row count. It leaves the existing free helper unchanged; compaction remains top-level only. 3. The null-payload test now explicitly releases the original source binding and bases ownership on allocation-start comparisons while the input is alive. I removed the misleading post-drop reread as an ownership argument. The sliced multibuffer test now selects `(1, 0)`, checks exact selected-only payload length/capacity, and excludes every source allocation, including sliced-away blocks. Ordinary interleave is a negative control demonstrating source sharing. Disassembly while investigating the benchmark controls also showed two additional outlined setup calls on the existing path. I restored the original free function body and added an inlining hint to `Interleave::new`; those calls are absent in the measured build. The description reports only exact-base/current-head results, including adverse cases and control limitations, without claiming universal speedup or zero overhead. Validation: 462 library tests and 21 doc tests with forced array validation, Clippy with warnings denied, formatting, and all 168 benchmark smoke cases passed. New batch tests cover all option combinations, schema/field metadata, mismatched and empty inputs, zero-column row counts/bounds, and nested views retaining ordinary behavior. I refreshed the evidence in the description and the existing reproducibility comments. This remains a draft; the earlier API-placement/coalescer discussion is still open. -- 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]
