Failed build on master branch (tests #202)

[github-actions]

Workflow failed: tests #202

[jayqi]

So what seems to have happened:

• Something in the last few days added certifi as a dependency in a fresh Python 3.6 conda environment.
  • This build on Python 3.6 with certifi-2016.9.26 installed into the fresh test environment
  • Last successful master build on Python 3.6 doesn't have it
  • This build on Python 3.8 doesn't have it
• Later, pip tries to upgrade certifi to certifi-2021.10.8 but runs into trouble uninstalling the old version that was installed by conda and not pip. This does seem like it's necessary because something in our dependencies requires certifi>=2017.4.17

[jayqi]

So it's installing this old version of certifi because for some reason, the conda resolver is installing this old version of setuptools for Python 3.6.

• This build: setuptools-49.6.0
• Last successful master build: setuptools-58.0.4

I'm actually even able to reproduce this on my laptop which is osx-64 vs. the build's linux-64. The conda resolver does this for some reason, but the mamba resolver does not.

[jayqi]

Well, this is a little mysterious, and I'm not sure what the best way to fix it is:

• Locally, running conda update setuptools doesn't do anything.
• Installing anything else that needs a later version of certifi updates it without the setuptools complaining. Newer versions of certifi don't have this uninstallation error, so pip can uninstall them okay.

I tried using mamba and mambaforge but something about mamba installations breaks on Windows.

Seems to suggest that something weird is going on between the conda resolver and the current conda metadata. We could maybe hard-code something to update certifi but that seems annoying to have baked into our CI. I'd prefer waiting to see if maybe it will just fix itself in a day or two.

[jayqi]

I've opened an issue on the conda-forge Python feedstock. conda-forge/python-feedstock#515