How to Handle 0 values on log scales

ref: https://github.com/bokeh/bokeh/issues/6517 ref: https://github.com/bokeh/bokeh/issues/8061

Here is the screenshot of 1D histogram, drawn with multi_line and y_axis_type="log" in bokeh==0.12.4. Background was done using plot.patch

Same plot using 0.12.6:

Also in 0.12.6, .patch is used for plotting blue background

[erickee]

Numpy handles 0 values in log axes properly. They (conceptually) map to -inf, which has the following implications

1. Line plot example:
   x = [1, 2, 3], y = [1, 0, 2] would become x = [1, 1, 3, 3], y = [1, -ymin, -ymin, 2] where ymin is a value that must always be below the bottom of the y axis limits, ideally even when zooming out.

2. Bar plot. Zero bars are conceptually the same as the line plot, just applied to bars.

This is really valuable for plotting probability distributions and histograms that have zero bins, and I hope Bokeh will match something like the behavior of numpy/pyplot.

[bryevdv]

• I am not sure anything uses (single value) compute, it’s there for completeness but I can’t say if anything uses it for sure.

• IIRC in our usage source is data space and target is screen