GeezerVison

This looks good. I assume the coverage reduction is a false positive due to the smaller number of total lines, due in turn to code removed from setup.py?

FWIW for anyone else reviewing, setuptools_scm looks like a safe, well-maintained dependency, with 18M downloads per month at time of writing.

Some remarks:

• should we specify setuptools >= 45 ? (see below)
• should importlib_metadata be in pyproject.toml, or does setuptools_scm already declare a dependency on it for Python < 3.8?
• _skbuild tripped me up a few times; it's a pity we can't have multiple of these with each having, say, a venv-specific suffix (easier said than done, I imagine).

I've done a bit of testing; below is what I found on my Ubuntu 20.04 machine. Main issue is that conda in-place build fails.

python3 -m build works fine, executed as follows:

#!/bin/bash

# build with `python -m build`
set -e

outdir=$(mktemp -d -t slycotwheel-XXXXXX)
venvdir=$(mktemp -d -t slycotvenv-XXXXXX)
echo "Wheels to to ${outdir}; venvdir is ${venvdir}"

python3 -m build ~/src/slycot --outdir ${outdir}
python3 -m venv ${venvdir}
source ${venvdir}/bin/activate
pip install ${outdir}/*
python -c "import slycot; print(slycot.__version__)"
pip install scipy pytest
python -c "import slycot; slycot.test()"

A non-conda in-place build works, though the first time I tried it I got this message:

/tmp/slycotvenv-5mwTyB/lib/python3.8/site-packages/setuptools_scm/integration.py:27: RuntimeWarning: 
ERROR: setuptools==44.0.0 is used in combination with setuptools_scm>=6.x

Your build configuration is incomplete and previously worked by accident!
setuptools_scm requires setuptools>=45

I changed my venv creation to have a newer version of setuptools, and all was fine (see below).

#!/bin/bash
set -e

venvdir=$(mktemp -d -t slycotvenv-XXXXXX)
python3 -m venv ${venvdir}
echo "venvdir is ${venvdir}"
source ${venvdir}/bin/activate
pip install "setuptools>=45.0" "setuptools_scm>=6.3" "wheel" "scikit-build>=0.14.1" "cmake" "numpy!=1.23.0"
cd ~/src/slycot

python setup.py build_ext --inplace

However, a conda in-place build did not work; the build is fine, but import slycot results in

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/rory/src/slycot/slycot/__init__.py", line 52, in <module>
    __version__ = imv(__package__)
  File "/home/rory/miniconda3/envs/sd-2/lib/python3.10/importlib/metadata/__init__.py", line 984, in version
    return distribution(distribution_name).version
  File "/home/rory/miniconda3/envs/sd-2/lib/python3.10/importlib/metadata/__init__.py", line 957, in distribution
    return Distribution.from_name(distribution_name)
  File "/home/rory/miniconda3/envs/sd-2/lib/python3.10/importlib/metadata/__init__.py", line 548, in from_name
    raise PackageNotFoundError(name)
importlib.metadata.PackageNotFoundError: No package metadata was found for slycot

conda in-place build:

conda create -n sd-2 --channel conda-forge python numpy scipy pytest gfortran_linux-64 gxx_linux-64 scikit-build setuptools setuptools_scm importlib_metadata ninja
conda activate sd-2
git clean -xfd  # in working tree, obviously
conda develop .
python setup.py build_ext --inplace -- -DNumPy_INCLUDE_DIR=$(python -q -c 'import numpy; print(numpy.get_include())') -DCMAKE_Fortran_COMPILER=x86_64-conda_cos6-linux-gnu-gfortran -DCMAKE_C_COMPILER=x86_64-conda_cos6-linux-gnu-gcc

• should we specify setuptools >= 45 ? (see below)

Let's do that. setuptools_scm should do it themselves, but they only do it for their own build system, not in install_requires.

However, a conda in-place build did not work

This needs some further investigation. Maybe it is better to not rely on importlib[._]metadata and use a setuptools_scm generated version.py instead. Do we have any downstream packages relying on the slycot.version / slycot.version API?

Should we include such build schemes into the CI?

Thanks, that fixed the conda in-place build problem, and the other two builds are still good.

Do we have any downstream packages relying on the slycot.version / slycot.version API?

No idea. We have 6 dependents according to libraries.io. We're back to this, I reckon leave as-is.

Should we include such build schemes into the CI?

I'm not sure. The build setup that matters most to me for development is a conda environment with in-place build; that lets me easily test and debug Slycot on different Python versions. The build setup that matters most for use of Slycot is the conda-forge feedstock.

I think we can merge - did you want to change anything else?

Can we hold on for a little bit? I am considering to move more metadata from setup.py to pyproject.toml.