Dependencies

[CommonClimate]

Periodically we need to discuss issues surrounding dependencies, so this is the space to do it.

Latest installment: a recent failed CI test on github (but success on my machine) and some sleuthing by @alexkjames uncovered this uncomfortable truth: our environment got befouled by a conflict between the latest SciPy (>=1.11.0) and Cartopy. The conflict is known, but still awaiting resolution, and I don’t see any of us writing that fix.
Until it is done, my latest PR (#449, which now passes) has taken the drastic measure of forcing scipy=1.10.1. Needless to say, pinned versions are something we strive to avoid, but they are sometimes unavoidable, as in the present case. I encourage all of you to monitor the Cartopy issue and report back when you see a fix, so we can relax this pin. Thank you!

[fzhu2e]

After some tests, I realize that pip will try to install the latest version of the dependencies if not installed already, while conda will have some dependency check and fix processes. In our case, scipy is a dependency of kneed and statsmodels. If pip install pyleoclim, the latest version of scipy will be installed, causing the mentioned conflict. If we conda install statsmodels first (need to avoid -c conda-forge), scipy=1.10.1 will be installed, avoiding the mentioned issue.

Therefore, a quick fix on this issue is install statsmodels via conda first before pip install pyleoclim. Note that this is already done in CI.