[ENHANCEMENT]: Consider removing the use of `argument_type` and `result_type` from `hasher` in cuco `default_filter_policy`

[mhaseeb123]

Is your feature request related to a problem? Please describe.

The member type std::hash::argument_type is deprecated in C++17 and removed in C++20, so we should consider syncing with STL and remove them from our hashers as well as their use in bloom filter policies.

Ref: https://en.cppreference.com/w/cpp/utility/hash

Originally posted by @bdice in rapidsai/cudf#17289 (comment)

Describe the solution you'd like

Consider refactoring to not use hasher::argument_type and hasher::result_type across cuCollections.

Describe alternatives you've considered

Keep using the deprecated hasher type aliases until C++20

Additional context

No response

[PointKernel]

I thought about mentioning this during the code review but didn’t want to come across as nitpicking. Instead of

cuCollections/include/cuco/detail/bloom_filter/arrow_filter_policy.cuh

Lines 84 to 85 in b8429d4

	template <class Key, class XXHash64>
	class arrow_filter_policy {

we could make the hasher a template type:

template<class Key, template <typename> class Hash>
class arrow_filter_policy {
  ...
};

To use it, users can simply pass the hasher type without repeating the Key type again, e.g.

auto p = arrow_filter_policy<Key, cuco::xxhash_32>{...};

[mhaseeb123]

I thought about mentioning this during the code review but didn’t want to come across as nitpicking.

No worries about that. I am always for pursuing excellence and attention to the detail is how we all learn. 😄

I think this is a great idea. Let's do this. But we still would be using the member types (argument_type and result_type) from the hasher which Bradley mentioned

[PointKernel]

we still would be using the member types (argument_type and result_type) from the hasher

Half-half. The argument_type will simply be key_type (using key_type = Key). However, retaining result_type (should be hash_value_type of the policy) is still desirable. The STL simplifies hashers by always using std::size_t as the return type, but this approach is suboptimal, especially in terms of performance and for GPUs. Such oversimplification does not fit our goals and IMO should be avoided.

[mhaseeb123]

Such oversimplification does not fit our goals and IMO should be avoided.

Agreed. Let's make the suggested change to cuco::arrow_filter_policy and close this issue otherwise.

[mhaseeb123]

#654 removes these member types for the arrow_filter_policy.