Interestingly scatter doesn't appear to set the major tick labels?

import matplotlib
matplotlib.use('Qt5Agg')
import matplotlib.pyplot as plt

fig, ax = plt.subplots()

ax.scatter([1., 0.], [0., 3.])
for tt in ax.xaxis.get_majorticklabels():
    print(tt)

Returns:

Text(0,0,'')
Text(0,0,'')
Text(0,0,'')

Etc..

EDIT: Ooops, need to call fig.canvas.draw() to get the ticks updated. Learn something new every day...

Added a new test to test_categorical.py to test for this case

Also fixes #9494

Two birds with one stone is nice!

@story645 I don't quite know what you mean by "registered". The converter gets initialiazed and attached to the xaxis object. But subsequent calls don't get sent to __init__, they get sent to convert so convert has to be able to handle new categories. I think... If thats wrong, and there is a better solution, then by all means we should revisit it.

I'm uneasy about it all because I dont' quite understand why axes.plot worked in the first place. This doesn't break axes.plot, but its strange that it worked but others did not

OK, I don't like this at all:

scatter does _process_unit_info(xdata=x, ydata=y, kwargs=kwargs) which only does xaxis.update_units if the xaxis doesn't have units already:

in _base.py:

if xdata is not None:
            # we only need to update if there is nothing set yet.
            if not self.xaxis.have_units():
                self.xaxis.update_units(xdata)

So thats why scatter wasn't updapting for categoricals, because the axis already had its units assigned.

Plot, on the other hand, calls:

_process_plot_var_args():

which always calls _xy_from_xy(self, x, y) which I am pretty sure always calls: self.axes.xaxis.update_units(x) regardless of whether the units have been already set.

So, plot has significantly different units handling than other plotting functions; plot calls update_units no matter what, where as other functions that run through _process_unit_info(xdata=x, ydata=y, kwargs=kwargs) do not.

1. I don't like the inconsistency, and I somewhat think that plot is in the wrong here and that this could lead to issues like datetime being mixed with non-datetime trying to work and failing.
2. I think that what we are doing in this PR for categoricals is OK, but really just in the case of categoricals where the mapping is one-to-one rather than a real scaling.

I've also faced this inconsistency. When I tried to use an empty array for the conversions, this worked:

fig, ax = plt.subplots()
ax.plot(["!", "0"], [1, 2], ".")

but this didn't

fig, ax = plt.subplots()
ax.scatter(["!", "0"], [1, 2])

and in both cases the convertor yields array([ 0., 1.])