aggregate_fn: Return NaN instead of null for mean of all-null input

[dimitarvdimitrov]

Mean::finalize_scalar returned null when the count was zero, while the array finalize path computes sum/count = 0/0 = NaN for the same input. A mean over an all-null group therefore gave different results depending on which accumulator we're using.

vortex/vortex-array/src/aggregate_fn/fns/mean/mod.rs

Lines 85 to 93 in 90d7433

	fn finalize(&self, sum: ArrayRef, count: ArrayRef) -> VortexResult<ArrayRef> {
	let target = match sum.dtype() {
	DType::Decimal(..) => sum.dtype().with_nullability(Nullability::Nullable),
	_ => DType::Primitive(PType::F64, Nullability::Nullable),
	};
	let sum_cast = sum.cast(target.clone())?;
	let count_cast = count.cast(target)?;
	sum_cast.binary(count_cast, Operator::Div)
	}

Since nulls are skipped during accumulation (as in standard SQL aggregation), an all-null input is an empty mean. Both paths now let the division produce NaN.

Note this only matters for the count = 0 case: sum overflow still returns null.

[codspeed-hq]

Merging this PR will improve performance by 32.96%

⚠️ 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.

⚡ 3 improved benchmarks
✅ 1529 untouched benchmarks

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	chunked_bool_canonical_into[(1000, 10)]	35.4 µs	20.5 µs	+73.06%
⚡	Simulation	chunked_varbinview_into_canonical[(1000, 10)]	213.2 µs	177.1 µs	+20.42%
⚡	Simulation	chunked_varbinview_canonical_into[(100, 100)]	308.5 µs	273.5 µs	+12.8%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing mitko/aggregate/mean-all-null-nan (90d7433) with develop (eda4dd0)}

[joseph-isaacs]

This looks incorrect to me. I think this should return Null

[robert3005]

I thought that NaN made more semantic sense but you can argue either way. Just have to document what is mean of an empty array

[robert3005]

Basically I think NaN as a result composes from partials more naturally without any additional checks

[dimitarvdimitrov]

if we combine NaNs, then the associativity of combining partials breaks. mean([a, b, c]) != combine_partial(mean_partial([a, b]), mean_partial([c])) because NaN always turns anything else into a NaN, but we can ignore nulls.

both duckdb and datafusion return nulls instead of nans for all-null input.

opened #8389 to revert this commit and fix finalize