Make matplotlib an optional dependency

[esheehan-gsl]

We're using metpy for some of its interpolate functions in an AWS Lambda function to process some data and send it to a web frontend for visualization. Requiring matplotlib is just increasing the size of our container image and, I think, introducing some security vulnerabilities via ImageMagick (according to the container security scan). It would be really nice if there were a way to disable the visualization tools in metpy so we can use it just for its analytical tools.

[dopplershift]

I've had some thoughts like this before, it would be nice to streamline. I think everything should work just fine as long as metpy.plots isn't imported, so it's just a matter of setting the proper packaging metadata. We'd of course have to consider impacts on other users who expect things to work fully out of the box after installing.

Are you building your containers using pip or conda?

[esheehan-gsl]

I'm using pip.

[dopplershift]

I think you can achieve what you want today with:

pip install importlib_metadata importlib_resources numpy pandas pint pooch pyproj scipy xarray
pip install --no-deps metpy

Depending on what you're using, you may be able to drop even more of the first line.

I'd love to hear feedback on if that works or something breaks unexpectedly.

[esheehan-gsl]

Good tip. Thanks! I'll give it a shot.

[esheehan-gsl]

I think I got this working with

pip install numpy Pint pooch pyproj scipy traitlets xarray
pip install --no-deps metpy

But I discovered something else also depends on Matplotlib (I think Cartopy is the culprit). 🤦🏻‍♂️ 😁

[dopplershift]

I'm surprised traitlets was necessary since that should only be imported through something down in metpy.plots...

Not sure what's prompting you to install CartoPy--there's not a lot useful about it if you're not using Matplotlib...

[esheehan-gsl]

The origins of Cartopy in the image are silly and uninteresting, so I won't bore you with the details. I was using it for something I could do with pyproj directly, so it's been removed from the image.

I can run another test to see if I'm able to eliminate traitlets and let you know how it goes.

[dopplershift]

That would be helpful, thanks!

[esheehan-gsl]

Alright, I must've messed something up with Docker yesterday because now I'm getting an import error because it can't import matplotlib, even though my script doesn't import matplotlib.

[dopplershift]

Can you share the full traceback? It'd be good to see because I actually expected that to work.

[esheehan-gsl]

Looks like it's because I'm using metpy.units.

Traceback (most recent call last):
  File "/var/task/app.py", line 3, in <module>
    from metpy.interpolate import cross_section, log_interpolate_1d
  File "/usr/local/lib/python3.9/site-packages/metpy/__init__.py", line 35, in <module>
    from .xarray import *  # noqa: F401, F403, E402
  File "/usr/local/lib/python3.9/site-packages/metpy/xarray.py", line 31, in <module>
    from .units import DimensionalityError, UndefinedUnitError, units
  File "/usr/local/lib/python3.9/site-packages/metpy/units.py", line 253, in <module>
    units.setup_matplotlib()
  File "/usr/local/lib/python3.9/site-packages/pint/registry.py", line 2189, in setup_matplotlib
    from .matplotlib import setup_matplotlib_handlers
  File "/usr/local/lib/python3.9/site-packages/pint/matplotlib.py", line 12, in <module>
    import matplotlib.units
ModuleNotFoundError: No module named 'matplotlib'

[dopplershift]

That's good to know. We can consider moving that inside metpy.plots instead.

[samthurston]

I'd like to bump this issue, as package management/build systems such as poetry and uv don't have support for excluding nested upstream dependencies, it would be really helpful if the plotting functions were an extra. My use case doesn't seem to have any problem just removing matplotlib from the environment, but my application will not run with matplotlib installed. thanks!

the doc builds on my PR seem to be failing due to a problem with pandas doc server? I'm not sure how to trigger them to re-run