[C++][Compute] Provide a kernel property testing API

[asfimport]

See discussion #7410 (comment)

Testing against random data fuzzes kernel implementations and provides sanity checks across a wide swath of parameter space with minimal configuration. Currently our random tests have a lot of boilerplate for generating the inputs and a lot of ad-hoc code for computing the expected values. It might be worthwhile to have an interface for specifying randomized tests more uniformly.

Since kernels provide introspection of their input and output types we can generate inputs of those types (both scalar and array). For ScalarFunctions, a function with signature Result<std::shared_ptr<Scalar>>(const ScalarVector& args, const FunctionOptions*) will be sufficient to specify the expected behavior, and the expected output can be generated by applying that specification to the broadcast inputs.

Example impl 9085048

Reporter: Ben Kietzman / @bkietz

_{Note: This issue was originally created as ARROW-9135. Please see the migration documentation for further details.}

[asfimport]

Wes McKinney / @wesm:
If I have understood correctly, the implementation of property testing in PR 7410 requires that the developer provide a reimplementation the kernel logic in the unit test suite, is that right?

[asfimport]

Ben Kietzman / @bkietz:
That's accurate; currently other random tests do the same (for example

arrow/cpp/src/arrow/compute/kernels/aggregate_test.cc

Lines 53 to 95 in cd87ac9

	template <typename ArrowType>
	static SumResult<ArrowType> NaiveSumPartial(const Array& array) {
	using ArrayType = typename TypeTraits<ArrowType>::ArrayType;
	using ResultType = SumResult<ArrowType>;
	ResultType result;
	auto data = array.data();
	const auto& array_numeric = reinterpret_cast<const ArrayType&>(array);
	const auto values = array_numeric.raw_values();
	if (array.null_count() != 0) {
	internal::BitmapReader reader(array.null_bitmap_data(), array.offset(),
	array.length());
	for (int64_t i = 0; i < array.length(); i++) {
	if (reader.IsSet()) {
	result.first += values[i];
	result.second++;
	}
	reader.Next();
	}
	} else {
	for (int64_t i = 0; i < array.length(); i++) {
	result.first += values[i];
	result.second++;
	}
	}
	return result;
	}
	template <typename ArrowType>
	static Datum NaiveSum(const Array& array) {
	using SumType = typename FindAccumulatorType<ArrowType>::Type;
	using SumScalarType = typename TypeTraits<SumType>::ScalarType;
	auto result = NaiveSumPartial<ArrowType>(array);
	bool is_valid = result.second > 0;
	if (!is_valid) return Datum(std::make_shared<SumScalarType>());
	return Datum(std::make_shared<SumScalarType>(result.first));
	}

). It'd certainly be possible to instead rely on the kernel itself as in CheckScalarUnary, which defers validation of inputs which are all scalars.

[thisisnic]

This issue hasn't had activity in a long time. If it's still being worked on, please leave a comment. Otherwise, it will be closed on 23rd June.

Labelled Status: Stale-Warning for tracking.