Reconsider publishing on PyPI
I'm aware the issue was discussed in #327 3 years ago and PyPI was discarded "because many of graph-tool's C++ dependencies, e.g. boost, cannot be pulled in" (though old releases are there). But time as passed, and what's possible might have changed. After all, numpy, scipy, Qt5 are available on PyPI, while too being mostly python APIs to C, C++, Fortran. Similarly, conda-forge was discarded because "it requires one to package every single C++ dependency of graph-tool, for every possible combination of OS and architecture", yet we now have it. Judging by the number of downloads, this is much appreciated.
I understand well that it requires a fair amount of work to maintain wheels on PyPI, but I sincerely think that this is the single most important barrier to a greater adoption of graph-tool. It's a great library, and it's a pity to see users flock to NetworkX because pip install networkx
just works. In my experience (teaching networks to 100+ students), beginners care a lot about the easiest route to use the API, and only later about the rest (features, speed, API design). At that time, they are already fairly invested in say NetworkX, and graph-tool misses out. Beside direct users, binary distributions considerably simplify cloud deployments like CIs and https://mybinder.org. While conda was a great leap forward, PyPI would be the cherry on the cake. :)
I take this occasion to thank you again for the amazing work you put on this project. Looking forward to an eventual pip install graph-tool
!