Dear NumPy devs, one of our NumPy integration doctests in Cython has been failing on travis for a couple of weeks. I checked the changelog of CPython and Cython from the time when it started, but neither showed anything remotely related. However, the failing travis nodes use NumPy 1.14, and since there were PyPy related changes in that release, I suspect that this could be the cause. I only see the failures in Py3.5+, and they started in the CPython-dev integration test runs for py3.6-dev and py3.7-dev on travis, however those are built (NumPy is pre-installed in them). Here's an example of a failing run.

The now failing test is simply this:

    cdef int[:, :] array = create_array((4, 5), mode="c", use_callback=True)
    print(np.asarray(array)[::2, ::2])

Meaning, there is a buffer provider (not NumPy), we take a Cython memory view from it, and then wrap that with numpy.asarray() to print it. The difference is when the buffer provider will be deallocated. Previously, it was cleaned up immediately at function exit (and we print a confirmation in a cleanup callback when that happens). With the new NumPy version, it is no longer cleaned up immediately, but only at the next GC run. I added a conditional GC call to the test to trigger that and keep the test from failing (we didn't need a GC run before), but I thought I'd at least bring this change to your attention, in case it's unintentional.