Test is

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

time = pd.date_range('2000-01-01', periods=10)
depth = np.arange(20)
data = np.random.rand(20,10)

fig, ax = plt.subplots()
ax.pcolormesh(time, depth, data)
plt.show()

ping @QuLogic who made the original change. I think this is compatible w/ your change, but still lets us use DatetimeIndex....

@jklymak We have a helper to skip or import depending on if pandas is installed. I took the liberty of putting your test code in and pushed to your branch (just in the interest of expediency).

This seems to make sense in general, but I'm just wondering why convert_?units returns a list instead of an array-like type.

@QuLogic It appears convert_?units can accept list-like objects and doesn't do any type conversion. I'm not up enough on pandas objects to know what would happen if you tried to change the type, but I think a lot of the time-based tick labelling depends on the pandas object, so I'd be loathe to change it.

I think the DatetimeIndex object (the time sequence) is not recognized (it is not in the units registry), so it is being treated as a list of Timestamp objects (which are in the registry). I haven't been able to track down exactly how this is happening. Units handling is convoluted, to say the least, and all the more so when it is split between matplotlib and pandas, which is calling back into matplotlib. In any case, the pandas DatetimeIndex object fortunately has a ravel method which has the same effect is the values attribute, yielding an ndarray of dtype 'np.datetime64[ns]'. This dtype is handled by the same registered converter as the Timestamp would be, tseries.converter.DatetimeConverter.

[testing for the jupyter backport-bot that this is a likely candidate for using soon]

@meeseeksdev hello

I'm Mr. Meeseek, @tacaswell, Look at meee !

@meeseeksdev backport to v2.1.x