numpy.ndarray binary incompatibility error in environments with `numpy<1.20`

After upgrading sisl to 0.11.0 I can not import the package anymore.

The call: from sisl._indices import indices_le returns: ValueError: numpy.ndarray size changed, may indicate binary incompatibility. Expected 88 from C header, got 80 from PyObject

OS Fedora 31 Python 3.7.6 numpy 1.17.5 (and others, see below)

Tried to reinstall numpy (same version) as suggested on the web, but couldn’t solve the problem. I am bound to this old numpy version due to other packages requirement, but I also tried in a new environment (same python version) numpy 1.18 and 1.19. Got the same error. Version of numpy 1.20, on the contrary, works.

[zerothi]

Yes, you are correct, it is basically the same as you suggest.

The only doubt I have is: does oldest-supported-numpy consider that sisl only supports >=1.13 or can it try to use an even older version? Would it matter if it picked an older version? I.e., is the requirement for >=1.13 a build time requirement or just a requirement for usage (say because a specific method was introduced in 1.13)?

I think I just fixed this. sisl only has a build requirement for the oldest numpy version available for the current python version. And once runned it requires a specific version. I just committed f364af06be2aa1854d8205bcad24c84efc598db8 which removed my extraneous requirement in requires! That variable is only a build requirement.

[zerothi]

I have now fixed it and tested some cases.

I am pretty sure (fingers crossed) that this is resolved for any later releases.

@pfebrer please have a look at the latest commit message and the file developments/test_numpy_installs.sh which tests a few things. You’ll note that this one will fail for v0.11.0 (due to not having proper build-dependencies) but will work regardless in the latest commits.

There were some things that needed updating, but I think this is on the right track.

Again, for future reference, one should be able to install sisl without problems if you pre-install the dependencies, and then subsequently install sisl with the --no-build-isolation to forcefully use the currently used versions:

python3 -m pip install -r requirements.txt
python3 -m pip install --no-build-isolation sisl-0.11.0.tar.gz

Note that if users manually install older numpy versions that do not follow the uploaded wheels for numpy, this procedure will fail and one really has to use the --no-build-isolation option.

I’ll close this, please re-open if not fixed.