On Wed, Aug 7, 2019 at 6:18 PM Charles R Harris <charlesr.harris@gmail.com> wrote:

On Wed, Aug 7, 2019 at 7:10 PM Stephan Hoyer <shoyer@gmail.com> wrote:
On Wed, Aug 7, 2019 at 5:11 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:

On Mon, Aug 5, 2019 at 6:18 PM Stephan Hoyer <shoyer@gmail.com> wrote:
On Mon, Aug 5, 2019 at 2:48 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:
The NEP currently does not say who this is meant for. Would you expect libraries like SciPy to adopt it for example?

The NEP also (understandably) punts on the question of when something is a valid duck array. If you want this to be widely used, that will need an answer or at least some rough guidance though. For example, we would expect a duck array to have a mean() method, but probably not a ptp() method. A library author who wants to use np.duckarray() needs to know, because she can't test with all existing and future duck array implementations. 

I think this is covered in NEP-22 already.

It's not really. We discussed this briefly in the community call today, Peter said he will try to add some text.

We should not add new functions to NumPy without indicating who is supposed to use this, and what need it fills / problem it solves. It seems pretty clear to me that it's mostly aimed at library authors rather than end users. And also that mature libraries like SciPy may not immediately adopt it, because it's too fuzzy - so it's new libraries first, mature libraries after the dust has settled a bit (I think).

I totally agree -- we definitely should clarify this in the docstring and elsewhere in the docs. An example in the new doc page on "Writing custom array containers" (https://numpy.org/devdocs/user/basics.dispatch.html) would also probably be appropriate.
As discussed there, I don't think NumPy is in a good position to pronounce decisive APIs at this time. I would welcome efforts to try, but I don't think that's essential for now.

There's no need to pronounce a decisive API that fully covers duck array. Note that RNumPy is an attempt in that direction (not a full one, but way better than nothing). In the NEP/docs, at least saying something along the lines of "if you implement this, we recommend the following strategy: check if a function is present in Dask, CuPy and Sparse. If so, it's reasonable to expect any duck array to work here. If not, we suggest you indicate in your docstring what kinds of duck arrays are accepted, or what properties they need to have". That's a spec by implementation, which is less than ideal but better than saying nothing.

OK, I agree here as well -- some guidance is better than nothing.

Two other minor notes on this NEP, concerning naming:
1. We should have a brief note on why we settled on the name "duck array". Namely, as discussed in NEP-22, we don't love the "duck" jargon, but we couldn't come up with anything better since NumPy already uses "array like" and "any array" for different purposes.
2. The protocol should use *something* more clearly namespaced as NumPy specific than __duckarray__. All the other special protocols NumPy defines start with "__array_". That suggests either __array_duckarray__ (sounds a little redundant) or __numpy_duckarray__ (which I like the look of, but is a different from the existing protocols).

`__numpy_like__` ?

This could work, but I think we would also want to rename the NumPy function itself to either np.like or np.numpy_like. The later is a little redundant but definitely more self-descriptive than "duck array".

NumPy-Discussion mailing list