How does Python / pip handle conflicting transitive dependencies?

Question

Suppose I want to install package a which requires the packages b1 and b2. In turn, b1 requires c > 1.0.0 and b2 requires c < 1.0.0. So the requirements of b1 and b2 cannot be fulfilled at the same time with the same package.

In principle / other programming languages, this is not a problem. One could install two versions of c side by side and make sure that b1 uses another version than b2.

However, I'm not sure if pip can install two versions of the same package. My first question is: Can pip install two versions of one package?

My main question is how one actually can deal with that problem. The only ways I can imagine right now is to

1. fork b1 (or b2) and a version of c that works for the fork, and upload b1_forked and c_for_b1_forked to PyPI, or
2. Include the code of b1 (or b2) directly in my project

Both seem more problematic than necessary.

What I tried

>>> import natsort; print(natsort.__file___)
'/home/moose/.local/lib/python3.6/site-packages/natsort/__init__.py'

$ cd /home/moose/.local/lib/python3.6/site-packages
$ ls
[... a lot of *.dist-info directories, some .py files, some .so files, ]
[... some directories called like the packages I've installed]

So I'm pretty sure this is where Python looks for installed packages and that only one version is installed (although the *-dist-info directories confuse me a bit).

This blog post suggests that there is no good solution for conflicting transitive dependencies at the moment. Do other projects (e.g. poetry) help with that?

Answer 1

In principle / other programming languages, this is not a problem. One could install two versions of c side by side and make sure that b1 uses another version than b2.

That's not a solution. If c manages a shared resource (console, e.g.) sooner or later b1 and b2 will stomp each other input or output via different cs and you end up with incorrect input and garbage output.

What you describe is a general problem, not limited to Python or pip. The only solution is to change b1 and/or b2 to agree on a version of c. Either downgrade b1 to allow c < 1.0 or upgrade b2 to allow c > 1.0.

Can pip install two versions of one package?

No, and the problem is not in pip but in Python: its import system doesn't allow importing from different versions of the same package. You can look at mitsuhiko/multiversion (Python2-only).