setup.py imports numpy, but numpy is an install time dependency

[mrshannon]

The setup.py file imports numpy which is an installation dependency. Because installation order cannot be specified (in order to ensure numpy is installed first), any package that depends on cf-units may also fail to install because numpy may not exist.

To reproduce the problem create a fresh virtualenv and run pip install cf-units which will result in:

pip install cf-units
Collecting cf-units
  Using cached https://files.pythonhosted.org/packages/ae/31/d803fd21a960db0cd9d03fd134e3551dfc96a8536956d8309caa4f44db90/cf-units-2.0.2.tar.gz
    Complete output from command python setup.py egg_info:
    Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/tmp/pip-build-1f2to47t/cf-units/setup.py", line 7, in <module>
        import numpy as np
    ModuleNotFoundError: No module named 'numpy'
    
    ----------------------------------------
Command "python setup.py egg_info" failed with error code 1 in /tmp/pip-build-1f2to47t/cf-units/

Therefore, cf-units is effectively broken for the standard python tool chain.

[pelson]

Thanks for the report @mrshannon. Truth be told, the standard python tool chain is broken when it comes to depending upon numpy - there is no universal way to define setup.py time dependencies that works well today (e.g. https://stackoverflow.com/questions/54117786/add-numpy-get-include-argument-to-setuptools-without-preinstalled-numpy). There is promise that this will change with the recent introduction of pyproject.toml in PEP518, but even this appears to have some pretty terrible warts that prohibit uptake.

Given cf-units' dependence on a non-python binary (UDUNITS-2) it seems to me that you are already outside the realms of "standard python tool chain" (you are never going to be able to install cf-units from pypi unless we statically ship UDUNITS-2 with it, which I personally think would be a huge maintenance burden - and the wrong solution).

I know it isn't to everybody's tase, but we have worked hard to make cf-units installable through conda conda install -c conda-forge cf-units as well as making it relatively easy to pip install once you have UDUNITS-2 and numpy installed (I recently removed the install-time dependence on Cython for source distributions, but not when cloning from github).

In short, I don't think there is much to be done here. Do you have an alternative suggestion @mrshannon?

[mrshannon]

The method in the SO answer that @pelson mentioned does work for cf-units as shown in #145.

As for needing a non-python binary, that is not unusual. For instance lxml (a popular XML library) requires libxml2 to be installed on the system and installs easily from PyPI.

Needing to preinstall numpy would not be an issue if I where the final user, but the issue makes building packages that depend on cf-units hard to justify.

[pelson]

Closed by #145. Thanks for putting the effort in to see this resolved @mrshannon. 👍