Is there a good test for this?

Sorry, changed my mind about where to target this to again.

To be honest, looking at it again, I am not even that certain that that the constructor of that Shadow class makes a lot of sense. Specifically, if I got it correctly, the ox and oy values are first multiplied by the dots-per-point conversion factor (e.g. 100/72 at 100dpi), then interpreted as offsets indata coordinates.

Which "of course(???...)" means that the shadow can move when you save the figure (as savefig.dpi does not have to equal figure.dpi)

Looking at the history this goes back to JDH in 2004 so it pre-dates the transform machinery and I would guess added as a gee-whiz feature.

If you think it could be cleaned up for 2.2 👍

I'm not really convinced of the utility of providing this class, but heh.

A problem with interpreting ox, oy as offsets in data units is that the only two public APIs that rely on that class (other than Shadow itself) arepie(..., shadow=True) andlegend(..., shadow=True), and for pie charts (which set the axes to polar) you'd betternot interpret ox, oy in data coordinates.

So perhaps ox, oy should be interpreted either as axes coordinates or figure coordinates -- don't have a clear preference right now.

Edit: I think I'll let@dstansby figure out how to handle the same problem with arrows and just copy his approach :-)

@dstansby test has been added (a while ago, but no worries)

@dstansby Would you like to re-review and merge if it is OK?

Seems ready to go in for 3.0 so labeling as RC.