Add fused compare and filter kernels for Delta

[joseph-isaacs]

Summary

Stacked on #10233, which makes the compressor choose Delta for metrics timestamps and counters. The westermo and lo2 queries then compare and filter Delta columns, and both operations currently decode the whole array first. These kernels work one 1024-value chunk at a time in stack buffers and never materialize the decoded array.

Changes

• decode_chunk in delta_decompress.rs decodes a single chunk: undelta, then untranspose.
• CompareKernel for Delta compares against a constant of the same ptype.
  • Each chunk is decoded and packed straight into the result bit words.
  • Signed types compare in the unsigned domain by flipping the sign bit.
• FilterKernel for Delta decodes only the chunks the mask touches.
  • It walks the mask 64 bits at a time: empty words are skipped, full words are bulk-copied, and partial words gather their set bits.
  • It declines, so the existing paths run, in two cases:
    • the mask is contiguous, which the filter executor already turns into a zero-copy slice;
    • the mask touches ≥ 90% of chunks at ≥ 15% density, where decoding everything and then bulk filtering wins.
• Both kernels are registered as execute-parent kernels in delta/vtable/kernels.rs.
• Tests cover:
  • all six operators against negative thresholds;
  • sliced arrays with nulls;
  • filter masks that are scattered, contiguous, in the last chunk, made of full words crossing a chunk boundary, or dense enough to fall back;
  • filtering a slice.

Kernel benchmarks

The benchmark uses 2^20 values in five shapes: i64 series-major timestamps, i64 jittered timestamps, a u32 counter, an i32 negative sawtooth, and i64 with nulls. Each shape is run full and sliced. The baseline decodes the whole array, then runs the same compare or filter. All results were checked equal to the baseline. Figures are geomeans, with minimum–maximum in brackets.

Operation	Mask density	Speedup
compare (Gt / Lte / Eq vs median)	–	1.69x (1.10–2.72x)
filter, one run	< 1%	496x
filter, one run	1–15%	74x
filter, clusters of 64	< 1%	70x
filter, clusters of 64	1–15%	3.8x
filter, uniform random	< 1%	5.0x
filter, uniform random	1–15%	1.6x
filter, strided	< 1%	5.3x
filter, strided	1–15%	1.8x
filter, any pattern	≥ 15%	~0.9x

In the ≥ 15% rows the kernel declines and the existing path runs. Running the same benchmark with the kernel switched off gives the same timings, so the gap to the baseline is not caused by this PR. Some of these rows show single-run variance of 0.5–1.4x.

Query benchmarks

These use the westermo and lo2 suites from #10214 and #10216, run on the Delta-recompressed files from #10233. Each cell is the Vortex time with this PR over the Vortex time on the same files without it: 5 iterations per query, on a 4-core cloud VM.

Suite	DataFusion	DuckDB
Westermo, geomean of 15 queries	1.09x faster	1.03x faster
LO2, geomean of 28 queries	1.21x faster	1.03x faster

Queries that changed by more than 30% were rerun with 15 iterations:

Query	Before	After
westermo Q01, DataFusion	87.8 ms	29.4 ms
westermo Q14, DuckDB	32.6 ms	16.6 ms
lo2 Q12, DataFusion	62.3 ms	24.4 ms
lo2 Q13, DataFusion	61.7 ms	19.9 ms
lo2 Q16, DataFusion	84.2 ms	35.4 ms
lo2 Q06, DuckDB	107.5 ms	82.3 ms

The apparent slowdowns in the first run, lo2 Q02, Q03 and Q15 on DataFusion and westermo Q01 on DuckDB, returned to their earlier times on the rerun. lo2 Q22 on DuckDB has a median of 101 ms against 86 ms before, over 30 iterations. Its p10 is 87 ms, so the change is within this VM's run-to-run variance and not established either way.

Checks run:

• cargo test -p vortex-fastlanes --lib -- delta: 89 passed.
• cargo clippy -p vortex-fastlanes -p vortex-btrblocks -p vortex-bench --all-targets --all-features -- -D warnings: clean, with clippy::manual_map allowed for an existing lint in vortex-buffer that this PR does not touch.
• cargo +nightly-2026-09-10 fmt --all --check: clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JXCBFkSfZNsuNNRowQfwdQ

[codspeed]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

✅ 2097 untouched benchmarks
⏩ 503 skipped benchmarks¹

---

_{Comparing ji/delta-kernels (e2f97a9) with ji/ts-compression (b2edd3c)}

Footnotes

1. 503 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. ↩