If the nightly wheels build cleanly we can close this.

That assumes that this was tested...

I am confused as to what the user-experienced problem is. From the cpython link, the following fails:

import ctypes, numpy
numpy.asarray(memoryview(ctypes.c_long(42)))

on an older python, whether I use HEAD or numpy 1.11.

Edit: The error changed from a ValueError to a RuntimeError

@eric-wieser could you clarify if the problem is a different exception type or something deeper?

@mattip: I think the case that matters is np.asarray(ctypes.c_long(42)). Previously that worked because we secretly added and used ctypes.c_long.__array_interface__. The referenced PR always prefers the buffer interface (ie, calling memoryview) to using __array_interface__

Turns out my WSL has python 2.7.12, so I'm all set to test this:

>>> sys.version_info
'2.7.12 (default, Dec  4 2017, 14:50:18) \n[GCC 5.4.0 20160609]'
>>> bad = ctypes.c_long(42)
>>> ctypes.sizeof(bad)   # this failure requires 64-bit long (ie, not windows native)
8

# 1.14
>>> np.asarray(bad)
RuntimeWarning: Item size computed from the PEP 3118 buffer format string does not match the actual item size.
array(42)

# master
>>> np.asarray(bad)
RuntimeError: Item size computed from the PEP 3118 buffer format string does not match the actual item size.

@eric-wieser We can close this now?

Not unless things have changed since I ran my test above (which shows it failing)

Hmm. I vote that we document the required versions, with a strong suggest that folks keep current. Given that the problems are fixed upstream that seems reasonable. What about python 3.5?

EDIT: I'm going to propose dropping 3.4 in the 1.16 release.

In case it changes your mind, ubuntu 16.04 ships with python 2.7.12, which exhibits the ctypes bug

I could try tweaking #10882 to have a special case for ctypes objects,

Sure, give it a shot, but don't spend a lot of time on it.

Reopening - this is a real issue for ndpointer, which provides:

• A valid but unsupported PEP3118 interface
• __array_interface__

After this change, we now reject the PEP3118 interface, and never check __array_interface__.

This causes toinsson/pyrealsense#82, which unfortunately they never raised upstream.

#12380 has made the error message even more cryptic, so we really ought to fix this for 1.16.

The follow-on issue #12421 was left open but tentatively declared "won't fix", does that mean this issue is also "won't fix", or at least can be removed from the 1.16 milestone?

Yeah, this can be removed from the milestone

I also stumbled on this: I have unrelated buffer interface (inherited from ctypes impl) and my custom array_interface that I'd like to be used

@eric-wieser we have dropped support for python<3.8, I think we can close this?

Agreed with Matti, seems like we can close it with old Python versions long EOL and unsupported.