Looking downstream - scipy imports argwhere then never uses it, and matplotlib uses it but only on a 1d array.

I suspect the only users actually affected by this will be the ones who have code that is currently broken on 0d arrays anyway.

If we're deprecating nonzero for this case, we probably need an implementation here that does not use it.

Lets maybe put this off to 1.18?

Don't think there's any way for me to write release notes if we do that. Opening #13707 to track solutions to that problem.

Updated with a release note

Failing:

File "/home/travis/build/numpy/numpy/builds/venv/lib/python3.6/site-packages/numpy/core/shape_base.py", line 438, in <module>

    _size = getattr(_nx.size, '__wrapped__', _nx.size)

AttributeError: module 'numpy.core.numeric' has no attribute 'size'

Yep, caught out by a rebase. Had to tweak the imports some more to avoid cycles.

Builds are failing:

../../builds/venv/lib/python3.6/site-packages/numpy/core/shape_base.py:531: in _atleast_nd

    return array(a, ndmin=ndim, copy=False, subok=True)

E   NameError: name 'array' is not defined

Looks all good to me. Lets give it a shot for 1.18. The quest for making 0D arrays not weird continues :).