comphead opened a new pull request, #23195: URL: https://github.com/apache/datafusion/pull/23195
## Which issue does this PR close? <!-- We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax. For example `Closes #123` indicates that this PR will close issue #123. --> - Followup https://github.com/apache/datafusion/pull/23192#discussion_r3476455650 ## Rationale for this change `arrow::buffer::OffsetBufferBuilder` is a thin wrapper around `Vec<O>` plus a `last_offset: usize` running counter; every `push_length(n)` does a `checked_add` on `usize` and a `usize_as(O)` conversion. For per-row loops with a known upfront row count, a direct `Vec<O>` that stores the running offset via `offsets[row] + O::usize_as(len)` can save measurable work in tight per-row loops — provided the offset push is a meaningful fraction of per-row cost. I swapped the pattern in all eight `OffsetBufferBuilder` call sites in the repo (`array_normalize`, `array_filter`, `remove`, `replace`, `array_add`, `utils::general_array_zip_with`, `array_scale`, `encoding::delegated_decode`), benchmarked the three sites that have criterion benches, and found the win is **not** uniform. ## What changes are included in this PR? Replace `OffsetBufferBuilder<O>` with `Vec<O>` (preinitialized with `O::zero()` and finalized with `OffsetBuffer::new(v.into())`) **only** in `datafusion/functions-nested/src/remove.rs`, where benches show clean wins with no regressions. The other seven sites are left on `OffsetBufferBuilder` — benches showed flat-to-regressing results, see below. ## Are these changes tested? Existing unit tests, doctests, and sqllogictests (`array_remove*`) pass unchanged. No new tests — refactor is functionally equivalent. ## Are there any user-facing changes? No. ## Benchmark results The biggest win is `array_remove` ### `array_remove` | Bench | size 10 | size 100 | size 500 | |---|---:|---:|---:| | `int64` | −0.2% | −1.0% | **−50.0%** | | `n_int64` | −0.7% | +0.05% | **−23.1%** | | `all_int64` | +0.2% | −1.8% | **−15.1%** | | `strings` | +3.8% | +0.6% | **−4.6%** | | `boolean` | −0.01% | +1.1% | +0.3% | | `fixed_size_binary` | −0.06% | **−20.0%** | **−2.6%** | | `int64_nested` | flat | flat | flat | For others its more like noise -- 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]
