@bjlittle "I can't identify the root cause for the change in numbers here... hence why I've introduced a minimum pin for cartopy in iris (after all, this is a major release update of proj)
We either accept this change "as is"... or dig deeper"
... I'm still in favour of creating an issue to cover this, if that seems like a pragmatic compromise. Then we can dedicate some time to digging a bit further to understand the root cause of the issue (if possible), but bank this PR and move on.
The failing test checks a few specific points from a regridding result
It's a nearest-neighbour regridding from an unstructured cube to a target grid defined in a gnomomic projection.
It seems that the nearest-neighbour regridded values are different, which implies that the coordinate projection points must have shifted.
Clearly the results shift when going from Cartopy 0.19 to 0.20, which means (probably) PROJ 7.2 --> Proj 8.1
Possibly the target points are just 'unlucky' in falling "just between" input locations, in which case adjusting the target grid slightly might enable this to give the same, stable results for both old+new versions
Or, possibly, there is a change (or even a definite problem) relating specifically to the Gnomonic projection
As explained in this comment
The failing test checks a few specific points from a regridding result It's a nearest-neighbour regridding from an unstructured cube to a target grid defined in a gnomomic projection. It seems that the nearest-neighbour regridded values are different, which implies that the coordinate projection points must have shifted.
For now, we have simply adopted new result values.
Clearly the results shift when going from Cartopy 0.19 to 0.20, which means (probably) PROJ 7.2 --> Proj 8.1