On Sun, Nov 4, 2018 at 8:45 AM Marten van Kerkwijk <m.h.vankerkwijk@gmail.com> wrote:
Does the above make sense? I realize that the same would be true for `__array_ufunc__`, though there the situation is slightly trickier since it is not as easy to bypass any further override checks. Nevertheless, it does seem like it would be correct to do the same there. (And if we agree this is the case, I'd quite happily implement it -- with the merger of multiarray and umath it has become much easier to do.)

Marten actually implemented a draft version of this already in https://github.com/numpy/numpy/pull/12328 :). I found reading over the PR helpful for understand this proposal.

I guess the practical import of this change is that it makes it (much?) easier to write __array_function__ for ndarray subclasses: if there's a function where NumPy's default function works fine, you don't need to bother with returning anything other than NotImplemented from __array_function__. It's sort of like NotImplementedButCoercible, but only for ndarray subclasses.

One minor downside is that this might make it harder to eventually deprecate and/or contemplate removing checks for 'mean' methods in functions like np.mean(), because __array_function__ implementers might still be relying on this.

But so far, I think this makes sense.

The PR includes additional changes to np.core.overrides, but I'm not sure if those are actually required here (or rather only possible due to this change). I guess they are needed if you want to be able to count on ndarray.__array_function__ being called after subclass __array_function__ methods.

I'm not sure I like this part: it means that ndarray.__array_function__ actually gets called when other arguments implement __array_function__. For interactions with objects that aren't ndarray subclasses this is entirely pointless and would unnecessarily slow things down, since ndarray._array_function__ will always return NotImplemented.