BUG/VIS: can't add a line with DatetimeIndex/datetime64 as x-values on timeseries plot

[jorisvandenbossche]

If you do a time series plot:

s = pd.Series(np.random.randn(100), index=pd.date_range('1970-01-02', periods=100, freq='T'))
fig, ax = plt.subplots()
s.plot(ax=ax)

and then want to add a line through matplotlib using the index or datetime64 values, this does not work (it plots the data at a totally wrong place outside of sight in this case):

ax.plot(s.index, s.values, color='g')
# or
ax.plot(s.index.values, s.values, color='r')

This is because PeriodConverter cannot deal with datetime64 values (Period/PeriodIndex also cannot deal with them), while when plotting this through the pandas plot interface, the datetime64 values will first be converted to periods.

[jorisvandenbossche]

OK, I thought this was an easy fix (#18468), but apparently it is not because matplotlib does not respect the ax.xaxis.converter.

So the first s.plot(ax=ax) sets ax.xaxis.converter to PeriodConverter. I thought that letting PeriodConverter handle datetime64 data this would solve the above issue, but when doing a ax.plot(..) apparenlty does not use the set converter, but overwrites it with a new converter:

In [16]: s = pd.Series(np.random.randn(100), index=pd.date_range('1970-01-02', periods=100, freq='T'))
    ...: fig, ax = plt.subplots()
    ...: s.plot(ax=ax)
Out[16]: <matplotlib.axes._subplots.AxesSubplot at 0x7f326f52def0>

In [17]: ax.xaxis.converter
Out[17]: <pandas.plotting._converter.PeriodConverter at 0x7f3275a03c88>

In [19]: ax.plot(s.index, s.values, color='g')
Out[19]: [<matplotlib.lines.Line2D at 0x7f326bca04a8>]

In [20]: ax.xaxis.converter
Out[20]: <pandas.plotting._converter.DatetimeConverter at 0x7f3275a03c50>   <--- changed !!

Given that, this is not something we can solve in pandas.

I opened a PR with the small fix and xfailing test, as the actual fix is useful anyway IMO (#18468)