[connortsui20]

Summary

• Tracking Issue: Tracking Issue: Fix Canonical Decimal #10701
• Stacked on Execute interleaved decimal arrays #10940.

A decimal dtype's canonical width can be much wider than its values need. Decimal Widen stores them in a lower-precision child, so compatible operations can avoid decoding the parent-width buffer.

Changes

Choose the child precision from valid values and scale. Push comparisons, bounds, fitting fills and ordered aggregates to the child where they preserve the canonical result. Decimal sums keep the parent accumulator precision. Execution copies decimal values directly, so a widening cast that returns another Widen cannot create an execution loop.

Widen stays in memory and outside ArraySession. The default writer already executes it through its pipeline, so CD4 adds no serialization lowering. Round-trip tests cover that path, alongside kernel, fallback and conformance tests. Benchmarks compare canonical and Widen inputs across three precision pairs.

API Changes

Casting a DecimalArray to a higher precision at the same scale returns Widen when the canonical width grows. Add WidenArray::encode_decimal and the shared smallest_decimal_dtype helper.

[codspeed]

Merging this PR will degrade performance by 27.52%

2 warnings

⚠️ Walltime results may be unreliable

Some benchmarks used the Walltime instrument outside CodSpeed Macro Runners, which can make the results inconsistent.

Learn more about Macro Runners

See affected benchmarks

⚠️ 4 benchmarks are dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

⚡ 6 improved benchmarks
❌ 7 regressed benchmarks
✅ 2261 untouched benchmarks
🆕 42 new benchmarks
⏩ 370 skipped benchmarks¹
🗄️ 30 archived benchmarks run²

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	CPU Simulation	filter_powerlaw_by_random[10000]	670 ns	31,971 ns	-97.9%
❌	Walltime	mul_u64_nonnull_neon	20.1 µs	40 µs	-49.79%
❌	CPU Simulation	bench_compare_varbin[(50000, 2)]	133.8 µs	162 µs	-17.37%
❌	CPU Simulation	bench_compare_varbin[(50000, 64)]	135.6 µs	163.9 µs	-17.31%
❌	CPU Simulation	bench_compare_sliced_dict_varbinview[(7500, 10000)]	200.8 µs	227.7 µs	-11.78%
❌	CPU Simulation	bench_compare_sliced_dict_varbinview[(10000, 10000)]	207 µs	233.8 µs	-11.47%
❌	CPU Simulation	bench_compare_sliced_dict_varbinview[(1000, 10000)]	79.9 µs	90.2 µs	-11.37%
⚡	CPU Simulation	widen_nullable[16384]	359.4 µs	224.8 µs	+59.88%
⚡	CPU Simulation	random_i8[0.01]	123.3 µs	98.9 µs	+24.66%
⚡	CPU Simulation	take[(16, 8, 1)]	119.5 µs	106.2 µs	+12.56%
⚡	CPU Simulation	max_f64	315.2 µs	281.2 µs	+12.07%
⚡	Walltime	add_constant_shapes_neon[(32768, PerRowConstant)]	15.7 µs	14.2 µs	+10.57%
⚡	Walltime	add_constant_shapes_neon[(32768, ConstantPerRow)]	15.7 µs	14.2 µs	+10.19%
🆕	CPU Simulation	cast_widen[p10_childp5_canonical]	N/A	913.8 µs	N/A
🆕	CPU Simulation	cast_widen[p10_childp5_widen]	N/A	789.7 µs	N/A
🆕	CPU Simulation	cast_widen[p18_childp9_canonical]	N/A	913.3 µs	N/A
🆕	CPU Simulation	cast_widen[p18_childp9_widen]	N/A	790.1 µs	N/A
🆕	CPU Simulation	cast_widen[p38_childp18_canonical]	N/A	1.8 ms	N/A
🆕	CPU Simulation	cast_widen[p38_childp18_widen]	N/A	1.5 ms	N/A
🆕	CPU Simulation	compare_constant_out_of_range[p10_childp5_canonical]	N/A	307.3 µs	N/A
...	...	...	...	...	...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

Tip

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

---

_{Comparing ct/decimal-widen (9577faf) with ct/decimal-interleave (c0eb004)}

Footnotes

1. 370 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. ↩

2. 30 benchmarks were run, but are now archived. If they were deleted in another branch, consider rebasing to remove them from the report. Instead if they were added back, click here to restore them. ↩