bisects to: #21989

Ah, so it seems that the change from super().autoscale(np.ma.array(A, mask=(A <= 0))) (which does produce min/max values for a fully masked array) to super().autoscale(in_trf_domain) (where in_trf_domain is a zero-length array) is the key change. I believe this would crash if log-scale is used any time all the bins are zero or negative.

I don't know what the correct solution is; is it as simple as putting a guard in autoscale like in autoscale_None that checks if A has a non-zero size? I don't know if this would have other bad downstream effects.

I suspect adding that guard is the right thing to do. It is impossible to correctly autoscale with no data so it is a bit of an undefined operation. In this case doing anything is better than crashing. It would be best if we still maintained the old behavior.

Putting in a guard (hacking my installation and not running a full test) does work in terms of not crashing, but leaves the hexbin (in particular) with no background rather than the default bluish background. I'm not sure if this is a problem, but it is different from the previous behavior (and the behavior without log scaling), but I don't know what the default values of vmin and vmax used to be in this case.

Closed by #23944.