DEP: Deprecate data_type.dtype if attribute is not already a dtype

[seberg]

Followup from gh-13003, marking as draft because that one needs to be merged (and then this one rebased) before it can go in.

This could hit the mailing list, but I think the deprecation here should be pretty uncontroversial, since it only enforces that something.dtype returns a dtype in statements such as np.dtype(something). It does not actually force the user to write the .dtype yet.

[BvB93]

A question: is it the plan to (eventually) move forward with this PR or is that still undecided?

[seberg]

Not sure why it stalled, unless anyone knows of a user who expects obj.dtype.dtype to be called, and even if, this is just a deprecation.

[eric-wieser]

Suggested change

	"`np.dtype(data_type.dtype)`. This will raise an error "
	"in NumPy 1.19.") < 0) {
	/* NumPy 1.19, 2020-02-07 */
	"`np.dtype(data_type.dtype)`.") < 0) {
	/* NumPy 1.20, 2020-10-19 */

[BvB93]

IMO it seems like a reasonable change and it makes typing a bit easier, so +1.

[mattip]

hmm. The circelCI merge-before-build command seems to be causing problems. Maybe we should allow it to fail silently.

[mattip]

The deprecation warning is leaking out into tests

[seberg]

@BvB93 its a new typing test that is failing (and not running locally with the slow flag). Should the test be modified or catch the warning?

[BvB93]

Should the test be modified or catch the warning?

Ah yes, the dtype value should be changed from float to something along the lines of np.dtype(float):

numpy/numpy/typing/tests/data/pass/dtype.py

Lines 31 to 32 in 136185c

	class Test:
	dtype = float

[seberg]

Thanks! Just wasn't sure which direction to fix it best.

[seberg]

@mattip I think the circleci is probably only because I had it set up on my branch. Disabled it now, it is not something I use regularly anyway.

[mattip]

Is this change required to avoid the deprecation warning? Does that mean all user code must now change?

[seberg]

No, don't remember why I did it here, but it is just maintanence, there must have been an issue with dtype being overshadowed somewhere.

[mattip]

Could you move the deprecation test to the proper file? Otherwise this looks good to me.

[seberg]

Done, we could consider deprecating even more, but I guess this might be a good (and easy) first step.

Yeah, I think I first wrote that at a time when I was considering to stop relying on that file so much (The original reason, and most of the magic in that file, was to work around python issues with warnings, which were eventually fixed).

[mattip]

Thanks @seberg

[charris]

@seberg I notice that this created changelog/13578.deprecation.rst in the root directory. What is that about? Should it be in the release notes?

[seberg]

Oh wow, that is a while ago. Yeah wrong place, should have been part of the release notes...

[charris]

OK, I'll make a PR to get rid of it in main and take care of it in the 1.21.x branch.