Cache min/max indices in patches

[palaska]

Cache min/max patch indices in Patches and short-circuit search_index when the query falls outside that range. Speeds up ALP/BitPacked scalar_at. This should help with slicing any patched array into a region that doesn't overlap the patch range

[codspeed-hq]

Merging this PR will degrade performance by 31.73%

⚠️ 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.

⚡ 16 improved benchmarks
❌ 71 regressed benchmarks
✅ 1119 untouched benchmarks

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	WallTime	cuda/bitpacked_u8/unpack/3bw[100M]	352.3 µs	298.6 µs	+17.96%
❌	Simulation	chunked_dict_primitive_canonical_into[f32, (1000, 100, 100)]	686.1 µs	766.3 µs	-10.46%
❌	Simulation	chunked_dict_primitive_canonical_into[u32, (1000, 100, 100)]	685.8 µs	764.6 µs	-10.31%
❌	Simulation	chunked_dict_primitive_canonical_into[u32, (1000, 10, 100)]	670.5 µs	749.3 µs	-10.52%
❌	Simulation	chunked_dict_primitive_canonical_into[u32, (1000, 100, 10)]	87.4 µs	97.2 µs	-10.08%
❌	Simulation	chunked_dict_primitive_canonical_into[f32, (1000, 10, 100)]	670.4 µs	750.6 µs	-10.69%
❌	Simulation	chunked_dict_primitive_into_canonical[f32, (1000, 10, 100)]	727.7 µs	808.6 µs	-10%
❌	Simulation	chunked_dict_primitive_into_canonical[u32, (1000, 100, 10)]	97.6 µs	108.7 µs	-10.22%
❌	Simulation	decode_primitives[f32, (1000, 2)]	17 µs	19.8 µs	-14.19%
❌	Simulation	decode_primitives[f32, (1000, 32)]	17.1 µs	19.3 µs	-11.85%
❌	Simulation	decode_primitives[f32, (1000, 4)]	17 µs	19.9 µs	-14.71%
❌	Simulation	decode_primitives[f32, (1000, 512)]	18.3 µs	20.9 µs	-12.34%
❌	Simulation	decode_primitives[f32, (1000, 8)]	17 µs	19.3 µs	-11.75%
❌	Simulation	decode_primitives[i64, (1000, 2)]	19.3 µs	22.3 µs	-13.65%
❌	Simulation	decode_primitives[i64, (1000, 4)]	19.3 µs	22 µs	-12.43%
❌	Simulation	decode_primitives[i64, (1000, 512)]	21.4 µs	24.3 µs	-11.91%
❌	Simulation	decode_primitives[i64, (1000, 8)]	19.3 µs	22 µs	-11.94%
❌	Simulation	decode_primitives[u8, (1000, 2)]	16.4 µs	18.8 µs	-12.67%
❌	Simulation	decode_primitives[u8, (1000, 32)]	16 µs	18.4 µs	-13.09%
❌	Simulation	decode_primitives[u8, (1000, 4)]	16 µs	18.4 µs	-13.25%
...	...	...	...	...	...

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

---

_{Comparing bp/minmax-index (8bc3d8a) with develop (f307edc)}

[joseph-isaacs]

Indices arrays have min/max stats these should be used instead.

We might not currently save or use this, but by these benchmarks we should.

If these benchmarks are cleaned up I am happy to merge those first.

[joseph-isaacs]

See: https://github.com/vortex-data/vortex/blob/e03bd1dcc63a34f260a16c05395146e26d1c9df3/docs/developer-guide/benchmarking.md

[robert3005]

Meta point, this is why you want to have special logic that's shared across all arrays, if we have Patched array then that could be one place where all of this logic could live without hacks. Alas we need Patched array to get over the line

[palaska]

Stats lookups (rwlock + hashmap + scalar conversion) were too expensive to be in the hot path (search_index is called in tight loops) so I still went with a OnceLock approach, combined with stats.

[palaska]

I think RwLock -> ArcSwap change made setting stats more expensive..

[joseph-isaacs]

It does look that way. I wonder if it made reading them faster?

[robert3005]

Github is failing me - remove extra Arc around arcswap tihs is not going to work, arcswap isn't clone