I think there is another simililar thing in the argument parsing used by normal ufunc calls (currently inside PyUfunc_GenericFunction

Removed that too

Note that __array_wrap__ uses context and it is documented

I worry that this perhaps should go through a deprecation cycle. I agree we don't need it though, since __array_ufunc__ should cover the same functionality.

Either way, this will need to hit the mailing list, in case there really are users out there. We should probably tag the commit API not MAINT

Mailing list link

Hmmm, I have been silly, right. I would be fine with not going through a deprecation here. But we can actually go through it, right? We can try with the context passed, and if that succeeds give a deprecation warning?

It is too bad, as it seems like just dead overhead to even build the context, but... should we go with a single release deprecation here?

adjusting documentation.

We should make any use of context in the c-api an error.

EDIT: Other then that, I think we were fine with going ahead with error/changing the API (in the 15.01.2020 triage meeting call)

Squashed to a single commit just in case we need to revert. Also actively check that the C-API use is always NULL, which exposed some calls in datetime_busday.c and datetime_busdaycal.c.

Thanks Matti, hopefully nobody will notice. If anyone does and finds this, please do comment, we are prepared to revert if necessary.

👋 breakage of rather old and hacky code in casadi project here
Deprecation cycle would have been nice.
Let's see if this __array_ufunc__ saves the day..

https://xkcd.com/1172/

If it's a big deal, I suppose we could change this to a DeprecationWarning for 1.19.1

it seems my check for using context is somehow failing, since the issue over in casadi is not that the call to __array__ is raising, rather that it used to have a context argument and now no longer does.

Well, we stopped providing context in some cases (specifically ufunc calls), it would be good to know if the context was actually useful here for some reason. The Deprecation will only be on the part of others providing a context manually...

What is the return np.array([np.nan]) for, and how come you need __array_ufunc__ to fix this (as opposed to actually clean it up a lot)? The reason we removed it, is that it never really worked. For example I am pretty sure the mentioned code will fail for np.add(a, b, out=a) even though it should be identical to a += b.