I'm not sure which way we should go with this. Pandas and xarray and maybe others autolabel when plot is called. But you almost always have to modify the labels to have actual name of the variable so utility of stating the units seems margins to me. Given that the units don't work properly under change of units I would consider just removing this feature of plot

Given the current MPL behaviour is to label axes, I think this inconsistency should be fixed to show the label, and changing this behaviour more widely is orthogonal to this bug.

😨 This shows me I understand little of the unit handling.

Feature

I agree with @jklymak that adding a unit label is of questionable value.

Furthermore, the current behavior is quite obscure:

1. It would be understandable if the label appeared through setting axs[0].yaxis.set_units(cm), but setting that or not does not affect the label.
2. Instead plotting data with that unit makes it magically appear - unless you've already explicitly set a ylabel before. IMHO this is too clever. "The user has not set a label and plots without labels are bad. Let's take any information we have to make a label."
3. It only works for Line2D-based data. Nether scatter, bar, stem of fill_between show this behavior.

Consistency

We definitely should make this consistent. Given (3) it's plot() that is the odd ball here. - But granted, weighted by use, it's more important than any single other plot type.

Recommendation

I strongly suggest to deprecate showing the label and get over this magic mess. At our core we want to be a simple understandable library with which you can configure your plot step by step. Note in particular that we do not auto-label plot('x', 'y', data=data) with 'x' and 'y', which would be much more reasonable than this unit stuff.

I broadly agree with the above comment, but again would say that a long term plan for this labelling is orthogonal to this issue, which is about fixing a specific bug and making the labelling consistent whether set_units() has been called or not.

The thing is that the consistency fix (=adding the label in more cases) is going in the opposite direction of the long-term plan (not having the label at all).

I think the further point is that plot is the outlier here. none of the other plotting methods label axes (that I know of)

3. It only works for Line2D-based data. Nether `scatter`, `bar`, `stem` of `fill_between` show this behavior.

I'm pretty sure this is incorrect - I just checked and it works for scatter, bar, stem, and fill_between... so given this I think the original report is a bug to be fixed with the current behaviour, and if we want new behaviour that's a new issue.

from basic_units import cm
import matplotlib.pyplot as plt
import numpy as np

cms = cm * np.arange(0, 10, 2)

fig, axs = plt.subplots(1, 4, tight_layout=True)

axs[0].stem(np.arange(5), cms)
axs[1].bar(np.arange(5), cms)
axs[2].scatter(np.arange(5), cms)
axs[3].fill_between(np.arange(5), cms)
plt.show()

Interesting. OTOH does it work with pcolormesh or contour? (Sorry away from computer)

Yep, works with both pcolormesh and contour

Well, then I agree it should be consistent... Sorry for not checking...

Sorry for letting this slip. I'll have a look again and respond soonish.

Oh, the issue is not that scatter() etc. does not add the unit, but that axs[1].yaxis.set_units(cm) makes the unit vanish.

This fix ended up breaking pandas so we reverted, re-opening this issues and moving to 3.8

Further discussion is available at #25219 as to how the original fix affected pandas. (just linking here a bit more explicitly)