Currently:

ERROR: packaging 20.1 has requirement pyparsing>=2.0.2, but you'll have pyparsing 2.0.1 which is incompatible.
ERROR: tox 3.14.4 has requirement six<2,>=1.14.0, but you'll have six 1.11.0 which is incompatible.
ERROR: pandas 1.0.1 has requirement python-dateutil>=2.6.1, but you'll have python-dateutil 2.1 which is incompatible.

This is not failing due to a backward compatibility policy in pip (pypa/pip#6969 (comment)). Nevertheless we should create a consistent build environment.

It seems like we need to replace pip with either poetry or dephell for build dependencies installation (or wait till July for pip with dependency resolver).

As long as the tests still pass, I'm fine with waiting for the pip improvements. Probably we're not using the parts of the libraries that have increased dependency requirements.

I am very 👎 on bringing one of the environment managers into Matplotlib. There is still too much thrashing in that space.

I am 👍 on bumping the min pyparsing to 2.0.3 (will open a PR, from reading the change log, 2.0.3 fixes a bunch of regressions from 2.0.2). We last touched this dependency in #8754 . I suspect that there is some critical bug fix in 2.0.2 (which is what made packaging pin there). Oddly, the 2.0.1 tag is gone from github, but the files are still up on pypi.

We don't use tox, so 🤷‍♀️ that it is un-happy.

We probably should pin the 3.6 pandas version back to a version that supports dateutil. We bumped the python-dateutil dep last in #9317 .

Our guidelines (https://matplotlib.org/devdocs/devel/min_dep_policy.html#python-dependencies) say to support at least 12 months, but only bump if we have to and I don't think we have any compelling reason to drag these versions forward.