Fully agree, some consitency would be nice

This seems to be a bug in that the data isn’t being converted to the right form early enough.

I actually think this is likely downstream in pandas. The units converter called is pandas.plotting._converter.DatetimeConverter, and it doesn't convert to matplotlib datenums. If you call ax.scatter(x=df.time.values, y=df.x.values) then it does convert the values to matplotlib datenums.

I'm going to close as needing a downstream fix, but I may be just passing the buck and not understanding something.

Yep, I guess the problem is more that ax.scatter(x=df.time, y=df.x) doesn't work either.

What is the type that is failing to be promoted?

I suspect that this is going to boil down to a datetime64 issue...

Again, if you call ax.scatter(x=df.time, y=df.x) you get the Pandas datetime converter: pandas.plotting._converter.DatetimeConverter.

As far as I can tell, the DatetimeConverter doesn't do any conversion on the Pandas series object:

Adding the following print statements to convert_units in axis.py

        ret = self.converter.convert(x, self.units, self)
        print(self.converter)
        print('x', x)
        print('\n x type', type(x))
        print('\nret', ret)
        print('\nret type', type(ret))

<pandas.plotting._converter.DatetimeConverter object at 0x1147f7d68>
x 0   2018-01-01
1   2018-01-02
2   2018-01-03
3   2018-01-04
4   2018-01-05
5   2018-01-06
6   2018-01-07
7   2018-01-08
8   2018-01-09
9   2018-01-10
Name: time, dtype: datetime64[ns]

 x type <class 'pandas.core.series.Series'>


ret 0   2018-01-01
1   2018-01-02
2   2018-01-03
3   2018-01-04
4   2018-01-05
5   2018-01-06
6   2018-01-07
7   2018-01-08
8   2018-01-09
9   2018-01-10
Name: time, dtype: datetime64[ns]

ret type <class 'pandas.core.series.Series'>

As you can see, the Pandas unit conversion doesn't do anything to the series.

Just to re-iterate the idea this is in panda's court; if we deregister the pandas converter this all works fine due to #10638. We did #10638 so that pandas didn't have to register their converter by default. I think a todo is to run a bunch of tests....

ping @TomAugspurger.

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

pd.plotting.deregister_matplotlib_converters()

dates = [datetime(2018,7,i) for i in range(1, 5)]
values = np.cumsum(np.random.rand(len(dates)))

df = pd.DataFrame({"dates":dates, "values" : values})

plt.plot(df["dates"],  df["values"])
plt.scatter(df["dates"],  df["values"])

plt.show()

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

pd.plotting.deregister_matplotlib_converters()

dates = [datetime(2018,7,i) for i in range(1, 5)]
values = np.cumsum(np.random.rand(len(dates)))

df = pd.DataFrame({"dates":dates, "values" : values})

plt.plot(df["dates"],  df["values"])
plt.scatter(df["dates"],  df["values"])
plt.plot("dates", "values", data=df)
plt.scatter(x="dates", y="values", data=df)

plt.show()

all works...

Sorry for the delay. Yeah, this looks to be a pandas issue

diff --git a/pandas/plotting/_converter.py b/pandas/plotting/_converter.py
index beebf84b8..cc835cd95 100644
--- a/pandas/plotting/_converter.py
+++ b/pandas/plotting/_converter.py
@@ -320,7 +320,11 @@ class DatetimeConverter(dates.DateConverter):
             return values
         elif isinstance(values, compat.string_types):
             return try_parse(values)
-        elif isinstance(values, (list, tuple, np.ndarray, Index)):
+        elif isinstance(values, (list, tuple, np.ndarray, Index, ABCSeries)):
+            if isinstance(values, ABCSeries):
+                # https://github.com/matplotlib/matplotlib/issues/11391
+                # Series was skipped. Convert to DatetimeIndex to get asi8
+                values = Index(values)
             if isinstance(values, Index):
                 values = values.values
             if not isinstance(values, np.ndarray):

I'll put that up as a pandas PR.

Awesome! I still think #11664 is the right thing to do from matplotlib's point of view, assuming pandas is still planning to not automatically register the pandas converters at import pandas. #11664 natively unpacks the pandas data type using np.asarray.

@TomAugspurger: What is panda's long-term plan vis-a-vis registering the matplotlib converter when pandas is imported?

I'm going to close this as fixed on panda's end....

Long-term, we'd like to not have import pandas also import matplotlib and register our converters. People will need to either make a plot with DataFrame.plot or Series.plot (which will lazily load matplotlib and register the converter), or expicility register the converter with pandas.plotting.register_matplotlib_converters().

Right now we're supposed to have a warning if people are implicitly relying on our converters, but that's apparently not working...

Thanks!, Ideally, I think matplotlib should be able to natively unpack panda's frames w/o pandas' converters, and I have such a PR submitted (#11664), but I wanted to make sure this was still pandas' long-term goal before pushing that PR forward. (i.e. there is not much point in making things work with the pandas' converters deregistered, if they are always going to be registered by import pandas.)

It seems this has now become Schrödinger's converter.
The code from the original post works now if you either register the converter manually or deregister the converter manually. Meaning either of

pd.plotting.deregister_matplotlib_converters()

or

pd.plotting.register_matplotlib_converters()

makes this work. It will throw a warning if you do neither of the above. Is this a semantic problem? I guess one wouldn't expect to get the correct output by performing a registration or a deregistration, while leaving it in a transient state if doing neither?

Last I checked, importing pandas still registers their converter. If you unregister it, you are using matplotlibs converter.