Thank you! I think this makes sense from what you described (and the results).

Can you please add a test for this? For example, a check_figures_equal-test for two of the cases that were earlier different, see the following for an example:

Thanks for taking a look, @oscargus

I've added a test test_hexbin_mincnt_behavior_upon_C_parameter and verified that indeed it does fail without the fix, and passes with the fix.

Anyone can merge on green CI

Thanks @krooijers and congrats on your first merged PR in Matplotlib. Hope to see you around!

@oscargus @ksunden This issue was discussed at length in #16865 and #21381. There's at least one issue with the fix there: this makes the reducer (defaulting to np.mean) be called for empty arrays too, and that triggers a warning ("Mean of empty slice") as soon as a cell is empty (e.g., hexbin(randn(1000), randn(1000), randn(1000))). You can easily imagine that there are reducers which would instead throw on empty inputs instead of returning nan, which would make things even worse. At the very least one should add an API change note stating that the reducer needs to be able to handle empty inputs (unless one always sets mincnt to 1), and suppress the warning in the case of np.mean.

As argued in the linked threads, I think we should switch the default reduce_C_function to np.sum (because it usually makes more sense than np.mean), and possibly also deprecate support for mincnt when C is not passed (because in that case mincnt is more simply and more generally expressed as vmin (plus possibly setting a custom "under" value on the colormap)).