Compatability with networkx is broken

[adam2392]

Compatability with networkx is broken

[adam2392]

I think numpy, pandas, scipy have all removed support for Python3.8 at this point on main. I think scikit-learn is in the process as well too.

Their general strategy is support older versions of X (e.g. Numpy) that have Python3.8 compatibility only for critical bug-fixes, but otw, all new features must be Python3.9+ compatible.

Since Python3.8 is technically still supported in version 0.10.x, and if v0.11.x is on a release schedule that overlaps w/ Python3.8 eol, it seems reasonable to deprecate python3.8 support in v0.11.xdev, while still maintaining support for 3.8 via v0.10.x and below. Up to you all! Just wanted to bring it some attention because this issue came up in pywhy-graphs, because I maintain extremely tight integration w/ networkx.

[emrekiciman]

@adam2392 Thanks for bringing up this issue and the PR to drop python 3.8 support.

[bloebp]

I would definitely wait for dropping 3.8 support and restrict the networkx version. Even 3.8 is not that commonly used yet. So, taking out that support would probably break the dependency for many people.

[adam2392]

I would definitely wait for dropping 3.8 support and restrict the networkx version. Even 3.8 is not that commonly used yet. So, taking out that support would probably break the dependency for many people.

I'm actually surprised by that because Python 3.7 lost support earlier this year already for a few months, so if a new bug/security-issue is found, it won't even be fixed, which is scary to me.

I'm just wondering along the lines: "is there a way to push our users towards a more healthy/less-risky usage of Python?"

[emrekiciman]

Yea, here's the official timeline of support for python versions. 3.8 is already in security-fix-only mode and is going to end-of-life in about 11 months.

https://devguide.python.org/versions/

I agree that keeping track with networkx and other major dependencies would be good. Do we know if Numpy, Pandas or other ML libraries that EconML and causal-learn depends on switching over on a schedule too?

[bloebp]

Ok, we might aim at 0.11 release in December.

[emrekiciman]

Now that we are past the 0.11 release, do we want to revisit python 3.8 support in main? @adam2392 @bloebp @amit-sharma

[bloebp]

Yes, we can drop 3.8 support for 0.12+