Obviously the test is a good idea! Not sure what to do about the failure though ;-)

Thanks for looking at this so quickly,@dstansby! To get around the images I thought you might be able to look at coordinates from a transform within the axes and compare them to the norm somehow (make a 0-1 axis and thentr = cbar.ax.transData + fig.transFigure.inverted() or something similar to back out where the tick(x) is versus the norm(x), but I think the issue is that these norms that fail don't have an inverse which defeats that route. Maybe you can think of a better way of doing that though to avoid the image test if you're inclined. But, I also think the image fine.

test failure seems real

Yep, they are. Not sure I have time in the next couple of days to fix, so if anyone else fancies investigating, fixing and pushing to my branch before then feel free.

I pushed to your branch but ci doesn’t seem to like it. Can your force push the change?

I figured out the force push

This is a situation where the Norm not having an accompanying scale is what is causing all the problems. The current fix is fine, but it basically special casesLogNorm andNormalize, and everything else is handled on a fake linear scale.

ForSymLogNorm andLogitNorm there are scales, but I don't see one forPowerNorm.

My suggestion would be for a bit of re-architecting here, and make each norm have a scale property that colorbar could consult rather than us having to guess the scale. That would also trigger folks who write Norms to create a corresponding scale so the colorbar can work with the a real y scale rather than a faked one.

I'm OK with this as is for a temporary fix - it passes CI except for a Pandas deprecation warning, which isn't this PR.

I also don't think the symmetric log plot is correct above - the linear scale is supposed to be one decade of the logarithmic scale, and its clearly not. It would be interesting to see if it ever was, but I doubt it. This will take some sitting down to figure out and may take me a few weeks as I have other deadlines.

I also don't think the symmetric log plot is correct above - the linear scale is supposed to be one decade of the logarithmic scale, and its clearly not. It would be interesting to see if it ever was, but I doubt it. This will take some sitting down to figure out and may take me a few weeks as I have other deadlines.

The SymLog code has been broken forever I think, and I have been working on it on and off recently. Hopefully I'll have something to show for it soon, thanks for opening the issue.

Given that@anntzer has a factory for turning scales into norms, that would really make our lives alot easier as we'd only need to make sure the Transform is correct, rather than rewriting one in the Norm. I think if we have a colornorm it should have a corresponding scale for the colorbar, so enforcing some regularity would be good.

Sorry I botched this up. I'll push this as a new PR with my apologies.