boost: can Homebrew break out boost-python?

[tdsmith]

I'm trying to test a homebrew-python formula that depends on Boost.Python and building boost --with-python today on my test VM took 3 hours and weakened my will to live. If I showed up with a boost-python formula that depended on (mercifully bottled) Boost and built only the Python parts, is it likely to be accepted into core?

[MikeMcQuaid]

So the problem with bottling these is that some people want to use Homebrew Python and some people want to use system Python. As the community doesn't seem able to pick one we don't build Boost's Python bindings for now by default.

[tdsmith]

Yes, understood about leaving Python out of the bottles. 👍

The Boost.Python Build HOWTO says:
"If building binaries from source, you might want to supply the --with-python argument to bjam (or the --with-libraries=python argument to configure), so only the Boost.Python binary will be built, rather than all the Boost binaries."

This makes me think, maybe erroneously, that building Boost.Python can be accomplished less painfully than (and separably from) building all of Boost. Probably I should try it before continuing the discussion.

[tdsmith]

It looks like this can work!

Boost has at least two contact points with Python:

• Boost.Python is a C++ library that makes it easier to write Python extensions in C++. It is not a set of Python bindings to other Boost libraries.
• Boost.MPI has some Python bindings.

Building Boost.Python separately from the rest of Boost is easy and takes only a couple minutes. Here is my proof of concept boost-python.rb, which is hacked down from boost.rb; here is a diff view.

My work-in-progress vpython formula, which uses Boost.Python and "normally" depends on boost --with-python, both builds and functions with the combination of bottled boost and this boost-python formula instead.

I'm not sure how to build the MPI bindings without building Boost.Python so that this doesn't conflict with boost --with-python (this could be as simple as just removing the four Boost.Python library files before linking, since they represent such a small fraction of Boost build time) but I think this is an encouraging start.

[MikeMcQuaid]

If you can make a slimmed down, working version of both formulae we'd review a PR. I think this is an interesting approach, certainly.

[NikolausDemmel]

As a user of Boost.Python I would very much welcome being able to use most of Boost from bottle and only have to build a separate Boost Python formula. 👍

[tdsmith]

Some notes:

• Boost.Python is reasonably popular. 5 formulas in homebrew-science, 6 in main homebrew, and 1 in homebrew-python can depend_on boost --with-python.
• Boost.MPI Python bindings are not popular and are not an obvious choice for new development. No formulas in homebrew-science, -python, or main mention boost, python, and MPI together. The MPI bindings depend on Boost.Python.
• Building boost --with-mpi --with-python --without-single doesn't currently install the Python MPI bindings. Nobody's complained.
• I think --with-python3 is wrong in the current boost formula; if specified, Boost.Python is always built only against python3. The flag was added two weeks ago in Option for boost::python to use python 3 #30590. I think there haven't been complaints because nothing depends on it yet. The approach we'll need to use is discussed at boost - provide option to pour from bottle? #21824 (comment).

I'll try and have a PR together and tested this weekend.

[jacknagel]

Closing in favor of the pull request.