Software versions

OS: Linux 5.4.6 Browser: Chrome 78.0.3904.108 (Official Build) (64-bit) Python: 3.8.1 JupyterLab: 1.2.4 Pandas: 0.25.3 Bokeh: 1.4.0

Issue

When plotting a vbar_stack, the entire row in the data source is ignored if the first input column contains a NaN in a pandas.DataFrame.

If a standard Python dict is used as a data source, the output is plotted as expected.

from bokeh.plotting import figure, show
from bokeh.models import ColumnDataSource

data = dict(index=[1, 2, 3, 4],
            a=[4, 5, None, 7],
            b=[9, None, 7, 6])

source = ColumnDataSource(data)

f = figure()
f.vbar_stack(x="index", stackers=["a", "b"], source=source,
             width=0.5, color=["red", "blue"])
show(f)

source = ColumnDataSource(pd.DataFrame(index=[1, 2, 3, 4],
                                       data=[dict(a=4, b=9),
                                             dict(a=5, b=None),
                                             dict(a=None, b=7),
                                             dict(a=7, b=6),
                                            ]))

f = figure()
f.vbar_stack(x="index", stackers=["a", "b"], source=source,
             width=0.5, color=["red", "blue"])
show(f)

Expected behavior with dict as data source

Unexpected behavior with DataFrame as data source

Mitigation

I currently let Pandas fill the NaNs with zeros.

Thanks for Bokeh, guys!