Install and packaging improvements todo list

After the latest round of install pain and the recent thread “stop supporting 'setup.py install'; start requiring 'pip install .' instead”, it’s time to start fixing a few things. Here’s a list (ticked items are addressed in gh-6895, which is not merged yet):

• unconditionally import setuptools (will fix gh-6551)
• error on python setup.py install. Provide a switch --force for python setup.py install. Mailing list discussion here not going to fly for now, see this email about issues
• update all install docs
• make it even clearer in big bold letters in the docs that scientific distributions or things like Homebrew are preferred over building from source or even pip.
• error on python setup.py upload (saying "use twine upload -s instead")
• error on python setup.py upload_docs
• error on python setup.py easy_install
• error on python setup.py test (saying "use python runtests.py instead")
• remove setupegg.py
• error on python setup.py clean (saying "use git clean -xdf (or -Xdf ...) instead")
• change python setup.py --help to first show numpy-specific stuff before setuptools help info
• figure out a better way to test the various methods of installing numpy
  • install from sdist
  • install via bdist_wheel
  • easy_install (is still triggered via setup_requires). Note, direct use of easy_install by users is not supported - meaning it may work but if it doesn’t then don’t expect support (use pip instead). This can be tested by installing a dummy package that uses setup_requires='numpy' on TravisCI.
  • the recommended pip install .
  • in-place builds (pip install -e, build_ext, develop)
  • builds with --config xxx build followed by pip install .
  • with and without virtualenv (can make a difference)
  • what else?
• when pip upgrade is released (should be soon, maybe under a different name, see https://github.com/pypa/pip/pull/3194 and https://github.com/pypa/pip/issues/59), officially change our mind and recommend the use of install_requires to packages depending on numpy.
• figure out what to do about setup_requires (pip won’t support that any time soon)
• fix up the metadata. now sdist overwrites “author” with “maintainer”. Consider using setup.cfg or a new static metadata format that may materialize soon.
• do a check on version numbering correctness (see [https://github.com/numpy/numpy/issues/6431#issuecomment-147112731) and make np.lib.NumpyVersion accessible in numpy.version
• figure out what to do with bdist_egg, bdist_rpm, bdist_mpkg, etc. RPM is probably a valid format to generate, the rest we may want to put behind a --force flag at least.
• do we need to provide a conda build recipe? --> EDIT: no, leave this to conda-forge

For interested people needing a bit of context, the Python Packaging User Guide is the authoritative doc about Python packaging in general.

[mattip]

We still support easy_install (and use it in doc/Makefile) even though it has been deprecated. Should we remove support for it?

[rgommers]

This is all basically done, let’s close this.