My impression was that the deprecation warning was only active when run with PyPy or NumPy was compiled with a special flag. Is that incorrect?

I think the example in c-info.howto-extend also needs to be marked with a Numpy version and a comment about PyPy compatibility. In fact, it might be helpful to use #if PYPY_VERSION in the code to show both traditional and new versions of the code. Something like

#ifdef PYPY_VERSION
/* PyPy compatible version for Numpy >= 1.14.0 */
...
#else
/* Traditional version using ... */

Using NPY_ARRAY_UPDATEIFCOPY will always emit a ``DeprecationWarning` when creating the array but never when deallocating it.

Using NPY_ARRAY_WRITEBACKIFCOPY or PyArray_SetWritebackIfCopy() will emit a DeprecationWarning when deallocating the array unless PyArray_ResolveWritebackIfCopy is called.

Only on PyPy (or with the flag) using PyArray_SetUpdateIfCopy will emit a warning DeprecationWarning but never when deallocating it.

Edit: correct flag names

Using NPY_UPDATEIFCOPY will always emit a ``DeprecationWarning` when creating the array but never when deallocating it.

This is the one that worries me as there is no fallback for code intended to work with earlier NumPy versions. SciPy uses it for instance. Also, the official macro name is NPY_ARRAY_UPDATEIFCOPY, the other alias is out of date (scipy uses both and needs a fix).

SciPy has already updated for the new behavior scipy/scipy#8150

I suppose we could leave it as is and deal with it later if there are complaints.

Thanks, sorry for not getting everyone on board earlier

No problem. Merged, thanks @mattip .

I may have been a bit quick here, shouldn't the example be using NPY_ARRAY_INOUT_ARRAY2?

good catch