Matplotlib 3.0 compatibility

[TomAugspurger]

I have some local failures with matplotlib 3.x only

========================================================================= FAILURES =========================================================================
______________________________________________________ TestTSPlot.test_matplotlib_scatter_datetime64 _______________________________________________________

self = <pandas.tests.plotting.test_datetimelike.TestTSPlot object at 0x118a34cf8>

    def test_matplotlib_scatter_datetime64(self):
        # https://github.com/matplotlib/matplotlib/issues/11391
        df = DataFrame(np.random.RandomState(0).rand(10, 2),
                       columns=["x", "y"])
        df["time"] = date_range("2018-01-01", periods=10, freq="D")
        fig, ax = self.plt.subplots()
        ax.scatter(x="time", y="y", data=df)
        fig.canvas.draw()
        label = ax.get_xticklabels()[0]
>       assert label.get_text() == '2017-12-12'
E       AssertionError: assert '2017-12-08' == '2017-12-12'
E         - 2017-12-08
E         ?         ^^
E         + 2017-12-12
E         ?         ^^

pandas/tests/plotting/test_datetimelike.py:1495: AssertionError
_________________________________________ TestDataFramePlots.test_if_scatterplot_colorbar_affects_xaxis_visibility _________________________________________

self = <pandas.tests.plotting.test_frame.TestDataFramePlots object at 0x11c340160>

    @pytest.mark.slow
    def test_if_scatterplot_colorbar_affects_xaxis_visibility(self):
        # addressing issue #10611, to ensure colobar does not
        # interfere with x-axis label and ticklabels with
        # ipython inline backend.
        random_array = np.random.random((1000, 3))
        df = pd.DataFrame(random_array,
                          columns=['A label', 'B label', 'C label'])

        ax1 = df.plot.scatter(x='A label', y='B label')
        ax2 = df.plot.scatter(x='A label', y='B label', c='C label')

        vis1 = [vis.get_visible() for vis in
                ax1.xaxis.get_minorticklabels()]
        vis2 = [vis.get_visible() for vis in
                ax2.xaxis.get_minorticklabels()]
>       assert vis1 == vis2
E       assert [] == [False, False, False, False, False, False, ...]
E         Right contains more items, first extra item: False
E         Use -v to get the full diff

pandas/tests/plotting/test_frame.py:1106: AssertionError
_________________________________________________ TestDataFramePlots.test_if_hexbin_xaxis_label_is_visible _________________________________________________

self = <pandas.tests.plotting.test_frame.TestDataFramePlots object at 0x11918ba58>

    @pytest.mark.slow
    def test_if_hexbin_xaxis_label_is_visible(self):
        # addressing issue #10678, to ensure colobar does not
        # interfere with x-axis label and ticklabels with
        # ipython inline backend.
        random_array = np.random.random((1000, 3))
        df = pd.DataFrame(random_array,
                          columns=['A label', 'B label', 'C label'])

        ax = df.plot.hexbin('A label', 'B label', gridsize=12)
>       assert all(vis.get_visible() for vis in
                   ax.xaxis.get_minorticklabels())
E       assert False
E        +  where False = all(<generator object TestDataFramePlots.test_if_hexbin_xaxis_label_is_visible.<locals>.<genexpr> at 0x11c69c990>)

pandas/tests/plotting/test_frame.py:1127: AssertionError

Our CI will start failing once these hit defaults.

[TomAugspurger]

Interestingly, I don't get the first failure from pytest pandas/tests/plotting/test_datetimelike.py::TestTSPlot::test_ts_plot_with_tz, but it does reproduce when running all the tests on that class (cc @WillAyd), so something is leaking between tests.

[TomAugspurger]

Got lucky and found it quickly:

pytest pandas/tests/plotting/test_datetimelike.py::TestTSPlot::test_irregular_datetime64_repr_bug \
 pandas/tests/plotting/test_datetimelike.py::TestTSPlot::test_matplotlib_scatter_datetime64

fails.

[TomAugspurger]

The test_if_scatterplot_colorbar_affects_xaxis_visibility and hexbin tests are failing because we implemented a hack to workaround an issue in matplotlib. That hack is no longer necessary with mil >= 3 (yay) and is actually breaking stuff, so we just need to not do it anymore.

[javadnoorb]

I participated in that implementation. Sorry if it was too hacky and caused issues. But skipping it seems to reopen #10611 in ipython notebook when I upgrade pandas to 0.24.0rc1.

As far as I remember testing things in inline backend was a bit of a hassle, and we eventually ended up writing the tests in the default backend. Because of that my guess is that resolution of #10611 is not tested properly.