Declare build dependency on Cython using pyproject.toml

[mkoeppe]

https://snarky.ca/what-the-heck-is-pyproject-toml/

[dave-doty]

@mkoeppe Did pip install ppsim fail for you? We supposedly have it configured to install cython as a dependency using setup.py, which reads the dependencies from requirements.txt.

I'm curious if it failed, or if you are just suggesting pyproject.toml as a superior alternative to specifying the cython dependency. I hadn't heard of pyproject.toml before. The difference with setup.py seems perhaps to be the "build" step, which is necessary for Linux/Mac systems. (We include a pre-compiled binary for Windows since it is less likely to have the C compiler necessary to build the cython code.)

[mkoeppe]

The problem with your setup.py is that it cannot even be run if Cython is not already installed because of https://github.com/UC-Davis-molecular-computing/ppsim/blob/main/setup.py#L2
This is a common problem, and is solved in modern Python packaging (PEP 517/518) using pyproject.toml.

Also, declaring cython as an install-requires is likely not needed for your package because it is only needed at build time of the package. You can see this in your wheels -- they are prebuilt but still pull in Cython when they are getting installed.

[dave-doty]

We may also have to do this with numpy, since it is also imported in setup.py. Not sure why; there are some strange path-manipulation lines it is used for:

ppsim/setup.py

Lines 7 to 8 in c29b749

	inc_path = np.get_include()
	lib_path = join(np.get_include(), '..', '..', 'random', 'lib')

[mkoeppe]

Yes

[UnHumbleBen]

@dave-doty @EricESeverson what are the steps/commands used upload to PyPI? I want to try to reproduce the issue in a test package so I can test out the installation after adding the pyproject.toml file and see if the Cython import issue is resolved

[EricESeverson]

We may also have to do this with numpy, since it is also imported in setup.py. Not sure why; there are some strange path-manipulation lines it is used for:

ppsim/setup.py

Lines 7 to 8 in c29b749

	inc_path = np.get_include()
	lib_path = join(np.get_include(), '..', '..', 'random', 'lib')

This was some weird wizardry required to get the cython code to be able to directly access the numpy random libraries. I really don't understand how it all works.

[EricESeverson]

@UnHumbleBen These are the steps I have written down in a text file

change version number
delete files in dist
conda activate base
python setup.py sdist bdist_wheel
conda activate py38
python setup.py sdist bdist_wheel
conda activate py39
python setup.py sdist bdist_wheel
twine upload --repository pypi dist/*
input pypi user information