This seems to be the right solution, but this is going to need a test. You can use the demo code in the original issue as the basis for the test.

Targeting for 3.3.1 as release critical, because it's a regression in 3.3.0 introduced via #17378.

Added a test. Maybe plotting 100 points was overkill thinking about it. Let me know if you want me to reduce the number of points.

Pre-fix

Post-fix

This looks pretty good. Quick question for the peanut gallery. Technically speaking, a Poly3DCollection is probably also impacted by this bug because it derives from PolyCollection, which derives from _CollectionWithSizes. Don't know of any real use-case though of specifying sizes for poly3d collections, though (or if it even works?).

Is an image test really required here? Can you not pull the sizes from the produced collection, or is this re-ordering draw-dependent?

If that's not possible, then I think we can still use a figure comparison test instead. Just compare with a plot using individual scatter calls.

@QuLogic The reordering is draw dependant so I couldn't get a standard test to work using get_sizes. I took your advice and changed to a figure comparison test with individual scatter calls.

@tomneep Can you squash this to one commit (or we can squash-merge it)? As it is the image will still be in the history (and we are trying to keep the repo from getting even bigger than it already is).

@tacaswell Makes sense. I've rebased and squashed into a single commit.

Thanks @tomneep!