`ufunc did not contain a loop with signature matching types` error with NumPy 1.22

For example, running:

$ python -m numba.runtests numba.tests.npyufunc -vf

results in:

test_basic_gufunc (numba.tests.npyufunc.test_ufuncbuilding.TestGUfuncBuilding) ... ok
test_basic_gufunc (numba.tests.npyufunc.test_ufuncbuilding.TestGUfuncBuildingJitDisabled) ... ok
test_basic_ufunc (numba.tests.npyufunc.test_ufuncbuilding.TestUfuncBuilding) ... ERROR

======================================================================
ERROR: test_basic_ufunc (numba.tests.npyufunc.test_ufuncbuilding.TestUfuncBuilding)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/gmarkall/numbadev/numba/numba/tests/npyufunc/test_ufuncbuilding.py", line 89, in test_basic_ufunc
    check(a)
  File "/home/gmarkall/numbadev/numba/numba/tests/npyufunc/test_ufuncbuilding.py", line 84, in check
    b = ufunc(a, a)
numpy.core._exceptions._UFuncNoLoopError: ufunc 'add' did not contain a loop with signature matching types (<class 'numpy.dtype[int32]'>, <class 'numpy.dtype[int32]'>) -> None

----------------------------------------------------------------------
Ran 3 tests in 0.300s

[esc]

Cool, will be nice to have the newest NumPy and numba working again soon 😃. I am a bit curious, do you always pin the numpy version because you want to ensure full compatibility across the NumPy API? Mainly, I am wondering if we can reduce the “downtime” where the newest NumPy isn’t supported by numba (at least on a fresh install) a bit in some simple way, e.g. by pre-release testing (giving us a month or so to fix up NumPy and up to a month head-start on numba changes when necessary).

Yes, I am the right person to talk to about this. We do want to start tracking NumPy closer to ensure comaptability with upcoming releases. We usually pin Numpy on our release branch to the latest working version at the time to prevent any unfortunate surprises if a Numpy release comes out half-way during our own releases cycle. I think it will be good idea however to discuss upgrading out “infrastructure” in it’s own ticket.

[seberg]

I think there’s probably two separate issues?

Yes, I guess so. I believe I had mentioned it before, aside from --pre releases, we do have NumPy dev wheels available: https://anaconda.org/scipy-wheels-nightly/numpy If you test against those “nightlies”, you probably want to say it is OK for that test to fail. But it would give a nice heads up about upcoming changes and allow you to float issues to NumPy quickly. (I guess testing numba is difficult, so I am not sure how easy this is.)