fix: backends should have defaults for user-facing operations

[agoose77]

The safe default for ak._backends.backend_of is to fail if the inputs don't have a well defined backend e.g. for lists. However, for user-facing operations, this is not correct; we want to assume NumPy for these types. A grep shows that only a few operations need this fix.

I haven't added any tests because we should really do this as part of a coverage pass for all of our operations.

---

📚 The documentation for this PR will be available at https://awkward-array.readthedocs.io/en/agoose77-fix-backend-default/ once Read the Docs has finished building 🔨

[codecov]

Codecov Report

Merging #1940 (6a6eb9b) into main (31ce657) will increase coverage by 0.02%.
The diff coverage is 96.29%.

Additional details and impacted files

Impacted Files	Coverage Δ
src/awkward/operations/ak_to_backend.py	100.00% <ø> (ø)
src/awkward/_backends.py	83.33% <85.71%> (+0.76%)	⬆️
src/awkward/operations/ak_backend.py	100.00% <100.00%> (+19.23%)	⬆️
src/awkward/operations/ak_cartesian.py	89.76% <100.00%> (+0.08%)	⬆️
src/awkward/operations/ak_concatenate.py	96.07% <100.00%> (+0.03%)	⬆️
src/awkward/operations/ak_fill_none.py	95.12% <100.00%> (+0.12%)	⬆️
src/awkward/operations/ak_run_lengths.py	90.90% <100.00%> (+0.13%)	⬆️
src/awkward/operations/ak_transform.py	83.05% <100.00%> (+0.29%)	⬆️
src/awkward/operations/ak_where.py	91.83% <100.00%> (+0.17%)	⬆️
src/awkward/highlevel.py	75.50% <0.00%> (+0.33%)	⬆️

[agoose77]

@jpivarski would you be happy for me to change ak.backend to throw an exception if the input arrays don't have a common backend? Right now it returns "mixed", which is not a valid backend name, and has less utility now that single arrays cannot be mixed.

[jpivarski]

So backend_of would normally raise an exception if it can't infer a backend without an explicit default=cpu?

awkward/src/awkward/_backends.py

Lines 176 to 182 in 31ce657

	nplike = nplike_of(*objects, default=None)
	if nplike is not None:
	return _backend_for_nplike(nplike)
	elif default is _UNSET:
	raise ak._errors.wrap_error(ValueError("could not find backend for", objects))
	else:
	return default

Okay, yeah, it does. So you need to provide default=cpu for any user-facing functions.

Looks good; go ahead and merge it!

[agoose77]

Yes. My call was that we really care about the backend, and due to interop it's very hard to tell if we get it wrong (JAX can adopt NumPy, CuPy can adopt NumPy). Better to fail loudly.

[agoose77]

@jpivarski would you mind confirming on #1940 (comment) ? I know you're busy with lots of things at the moment, but don't want to merge this without an OK!

[jpivarski]

Yeah, "mixed" no longer means anything and shouldn't be returned by anything. I'm in favor of doing things this way, though we've got to be careful to always provide a default (and it would always be cpu) in user-facing functions now.

It is better to fail loudly.

I hadn't noticed that this is a draft PR. If it's a draft because you still have a lot of user-facing functions to update, you don't need re-approval to merge it. I'm in favor of the principle.

[agoose77]

Thanks! I'm going to make two changes; removing "mixed" from our high level API, and to associate the backend name on the backend object so this code becomes more trivial.

[jpivarski]

Raising an error in mixed cases is right. We should not have any array with different backends in different nodes.

[agoose77]

OK, this is done. I've added a check to ensure that we throw a ValueError if users somehow encounter a typetracer