BUG: plotting with DatetimeIndex containing NaT

[jorisvandenbossche]

xref #8914

In [1]: %matplotlib
Using matplotlib backend: Qt4Agg

In [2]: df = pd.DataFrame({'date': pd.date_range('2016-01-01', periods=5), 'vals
': range(5)})

In [3]: df.loc[2, 'date'] = np.nan

In [4]: s = df.set_index('date')['vals']

In [5]: s
Out[5]:
date
2016-01-01    0
2016-01-02    1
NaT           2
2016-01-04    3
2016-01-05    4
Name: vals, dtype: int64

In [6]: ax = s.plot()

In [11]: ax.get_lines()[0].get_data()
Out[11]:
(array([datetime.datetime(2016, 1, 1, 0, 0),
        datetime.datetime(2016, 1, 2, 0, 0),
        datetime.datetime(2016, 1, 4, 0, 0),
        datetime.datetime(2016, 1, 5, 0, 0), NaT], dtype=object),
 array([0, 1, 3, 4, 2], dtype=int64))

In [44]: ax.get_lines()[0].get_xydata()
Out[44]:
array([[  7.35964000e+05,   0.00000000e+00],
       [  7.35965000e+05,   1.00000000e+00],
       [  7.35967000e+05,   3.00000000e+00],
       [  7.35968000e+05,   4.00000000e+00],
       [  6.12411009e+05,   2.00000000e+00]])

So this gives you a plot with one of the values in the year 1677, September 22, so the minimum possible Timestamp.

This is of course not correct, but there are two things:

• NaT gets converted to Timestamp.min
• the order of the values is changed, as the NaT is put at the end

There is a related issue about plt.plot(pd.NaT) erroring (#9253), but in this case it is pandas code that gives wrong results, so we have more control over this. I find it also strange that this does not error as in the pd.NaT case, but converts the NaT to Timestamp.min (but didn't look into detail)