@anntzer Updated re discussion in #9774. This now allows set_xlim etc, before data is plotted:

This works:

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.dates as mdates
import datetime

base = datetime.datetime.today()
time = [base - datetime.timedelta(days=x) for x in range(0, 3)]
data = np.random.rand( 3)

fig, ax = plt.subplots(tight_layout=True)
ax.set_ylim(time[0], time[1])
ax.plot( data, time)

This errors:

fig, ax = plt.subplots(tight_layout=True)
ax.set_ylim(0., 1.)
ax.plot( data, time)

This version doesn't lockout data that would be registered by new data, it just uses the original converter first registered for the axis. Its up to the converter to handle the new data.

For instance, if we call:

base = datetime.datetime(2017, 1, 1, 2, 3, 10)
time = [base + datetime.timedelta(days=x) for x in range(0, 3)]
ax.plot([1., 2.], [1., 3.])
ax.plot(time) 

the call will fail. But:

base = datetime.datetime(2017, 1, 1, 2, 3, 10)
time = [base + datetime.timedelta(days=x) for x in range(0, 3)]
ax.plot(time) 
ax.plot([1., 2.], np.array([1., 3.]) + 736330)

will succeed.

How does this work in a sharex/sharey situation? Would plots in ax2 pick up the convertors (and mappings) from ax1?

Good question, I'm not sure. Do they do that now? Its doable, but I didn't do anything to make it happen.

As we discussed after today's meeting: I think that any lockout machinery should apply only to converters, not to the native floats to which unitized quantities are converted for plotting. For example, in making a plot of categoricals, one should remain able to place lines or any other artists on the Axes using native floats. Similarly, one should remain able to plot Datetimes and mpl datenums on the same Axes. This should be possible regardless of the order in which artists are added.

attn @jrevans @TD22057

@efiring, just to rephrase: the converter would be locked to the axis. However, the converter could still choose to deal w/ floats in a self-consistent way. So, the datetime converter could still be set to allow floats through, however, it might perform its own checks on them.

What that would preclude however, is calling plot with a float first and then with a date later and expecting the date to work. To get around that, there should probably be a mechanism to manually set the converter to an axis.

I still think this is a good idea, but will close until some more thought goes into units support.