VarBin: build views straight into a VarBinViewBuilder

[robert3005]

Appending a VarBin array to a VarBinViewBuilder canonicalized it first. That built the views, wrapped them in a VarBinViewArray the builder immediately unwrapped, and then had append_varbinview_array rewrite every view to rebase its buffer index onto the builder's.

Number the buffer up front instead: varbin_decode_views takes the index the pushed buffer will land at, so the views come out already correct and the append is one view per row plus a buffer push. varbin_to_canonical now shares that helper, which also shrinks what match_each_integer_ptype! duplicates — the array construction no longer expands once per integer width.

A builder configured to compact still goes the canonical route. It chooses per buffer whether to keep, slice or rewrite it by measuring the finished views against the buffer, and push_buffer_and_adjusted_views takes buffers as they are, so the fast path would silently opt it out. VarBinViewBuilder::compacts_buffers is the new accessor that guards it.

Benchmarks

vortex-array/benches/chunk_array_builder.rs, fastest of 100 samples, 3 runs per side:

	benchmark (rows × chunks)	before	after	speedup
⚡	chunked_varbin_to_varbinview_builder 10×1000	313 µs	204 µs	1.53x
⚡	chunked_varbin_to_varbinview_builder 100×100	87.3 µs	69.5 µs	1.26x
⚡	chunked_varbin_to_varbinview_builder 1000×10	58.6 µs	54.8 µs	1.07x
⚡	chunked_varbin_opt_to_varbinview_builder 10×1000	378 µs	266 µs	1.42x
⚡	chunked_varbin_into_canonical 10×1000	357 µs	253 µs	1.41x

The last row is the control — it never touches the builder, so its gain comes from the shared-helper extraction alone. Relative to it, the append itself drops from 0.88x to 0.81x of a canonicalization. Most of the absolute win at high chunk counts is the smaller macro expansion; the rest is the skipped view rewrite.

Checks

• cargo nextest run -p vortex-array -p vortex-fsst -p vortex-zstd -p vortex-onpair -p vortex-mask — 3377 passed
• cargo clippy -p vortex-array --all-targets --all-features — clean
• cargo +nightly fmt --all
• New tests: append_varbin_to_varbinview_builder (interleaves VarBin appends with staged in-progress buffers and a VarBinView append, so buffer numbering is exercised) and append_varbin_to_a_compacting_builder_still_compacts.

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will degrade performance by 48.15%

❌ 6 regressed benchmarks
✅ 1836 untouched benchmarks
🆕 9 new benchmarks
⏩ 55 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	extend_from_array_non_zctl_overlapping[(10000, 8)]	713.2 µs	2,462.4 µs	-71.03%
❌	Simulation	extend_from_array_non_zctl_overlapping[(1000, 32)]	298 µs	893.4 µs	-66.64%
❌	Simulation	extend_from_array_zctl[(10000, 8)]	1.2 ms	2.2 ms	-45.43%
❌	Simulation	extend_from_array_zctl[(1000, 64)]	734.9 µs	1,201 µs	-38.81%
❌	Simulation	extend_from_array_non_zctl_overlapping[(1000, 8)]	301.3 µs	429.9 µs	-29.93%
❌	Simulation	extend_from_array_zctl[(1000, 8)]	345.5 µs	402.1 µs	-14.08%
🆕	Simulation	chunked_varbin_into_canonical[(10, 1000)]	N/A	2.5 ms	N/A
🆕	Simulation	chunked_varbin_into_canonical[(100, 100)]	N/A	643.6 µs	N/A
🆕	Simulation	chunked_varbin_into_canonical[(1000, 10)]	N/A	535.1 µs	N/A
🆕	Simulation	chunked_varbin_opt_to_varbinview_builder[(10, 1000)]	N/A	2.8 ms	N/A
🆕	Simulation	chunked_varbin_opt_to_varbinview_builder[(100, 100)]	N/A	663.5 µs	N/A
🆕	Simulation	chunked_varbin_opt_to_varbinview_builder[(1000, 10)]	N/A	456.3 µs	N/A
🆕	Simulation	chunked_varbin_to_varbinview_builder[(10, 1000)]	N/A	2.1 ms	N/A
🆕	Simulation	chunked_varbin_to_varbinview_builder[(100, 100)]	N/A	592.5 µs	N/A
🆕	Simulation	chunked_varbin_to_varbinview_builder[(1000, 10)]	N/A	520 µs	N/A

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing rk/varbin-to-varbinview-builder (d4aaedf) with rk/onpair-decode-into-builder (8fa6c4f)}

Footnotes

1. 55 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩