feat(array): scalar probes with one-off and repeated probe types

[joseph-isaacs]

Adds reusable scalar probes with lazy state at each child slot. Repeated traversal of the same slot path reuses its preparation, including validity. ArrayRef::execute_scalar and is_valid now go through a one-off probe with the same bounds and validity checks.

Public API

// One-off read: borrows the array, retains nothing.
let value = array.probe().execute_scalar(10, &mut ctx)?;

// Repeated reads share preparation until the probe is dropped.
let mut probe = array.repeated_probe();
let first = probe.execute_scalar(10, &mut ctx)?;
let nearby = probe.execute_scalar(11, &mut ctx)?;
let valid = probe.execute_is_valid(12, &mut ctx)?;

ArrayProbe<'a> is the borrowed reader, either Once(&ArrayRef) or Repeated(&mut RepeatedArrayProbe). RepeatedArrayProbe owns its array handle and can outlive the original handle; it keeps the encoding's state, its validity, and one retained probe per child slot, all created on first use and dropped with the probe.

Encoding contract

pub trait OperationsVTable<V: VTable> {
    type ProbeState: Default + 'static;

    fn probe_scalar(
        state: &mut ProbeState<'_, V>,
        index: usize,
        ctx: &mut ExecutionCtx,
    ) -> VortexResult<Scalar> {
        Self::scalar_at(state.array(), index, ctx)
    }
...
}

Breaks

Need to add a type ProbeState: Default + 'static; to the OperationsVTable

[codspeed]

Merging this PR will degrade performance by 15.59%

⚠️ 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.

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 4 improved benchmarks
❌ 7 regressed benchmarks
✅ 2187 untouched benchmarks
⏩ 218 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	words_gather_dispatch_avx512[1024]	9 ns	34 ns	-73.53%
❌	WallTime	arrow_checked_add_u32_neon[16384]	13.5 µs	20.3 µs	-33.42%
❌	Simulation	random_i8[0.5]	71.3 µs	94.9 µs	-24.81%
❌	Simulation	new_raw_prim_test_between[i32, 2048]	64.1 µs	77.6 µs	-17.43%
❌	Simulation	decompress[u64, (4000, 1024)]	71.5 µs	85.9 µs	-16.71%
❌	WallTime	filtered_sink_i64_neon[NineNullsInTen]	15.4 µs	17.4 µs	-11.42%
❌	WallTime	filtered_owned_i64_neon[NineNullsInTen]	15.3 µs	17 µs	-10.17%
⚡	Simulation	execute_scalar_struct_wide	503 µs	389.5 µs	+29.15%
⚡	Simulation	random_i16[0.8]	96.4 µs	78.5 µs	+22.83%
⚡	Simulation	execute_scalar_struct_simple	106.5 µs	89.6 µs	+18.8%
⚡	Simulation	allocate_drop_arrow[0]	456.9 ns	402.7 ns	+13.45%

Tip

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

---

_{Comparing ji/array-probe-api (22e0f50) with develop (b5f43ba)}

Footnotes

1. 218 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. ↩

[robert3005]

I will leave one comment instead of mentioning the lines. I think the name churn is unfortunate but I think it's fine. The If we have ProbeUsage::Once we should replace execute_scalar with probe usage once invocation. The null checks are split and duplicated right now and this replacement would force us to unify it now. I think it's right that you want to let implementation decide.

Not sure we need to hold arrayref in ProbeAccess::Once if we pass it to the function explicitly

[robert3005]

since you have this arrayview then ProbeAccess::Once state is weird but I could be missing something

[joseph-isaacs]

This was not ready to review yet

[robert3005]

this addresses all the comments I had. The only question I had is whether we want to defer validity check to the impl instead of having default BUT I think if we fix validity computation with this it's better