Unexpected behaviour when merging cubelist with constant NaN-valued scalar coord

[btrotta-bom]

Merge seems to treat scalar coordinates differently when they have value NaN. I am trying to merge a list of cubes, each of which has a scalar coordinate with the same value. If this value is not NaN, the merge works as expected: the resulting cube has a scalar coordinate identical to the one in the input cubes. But if the value is NaN, this coordinate becomes 2-dimensional in the result.

Example below:

import iris
import iris.cube
import numpy as np
from iris.coords import DimCoord, AuxCoord

data = np.random.random((3, 3))
dim_coords_and_dims = [(DimCoord(np.arange(3), long_name="coord1"), 0)]

# aux_coord is constant non-nan value, merged cube contains aux_coord as a scalar coord as expected
aux_coord = AuxCoord([0], long_name="aux_coord")
cubelist = iris.cube.CubeList()
for i in range(2):
    for j in range(2):
        aux_coords_and_dims = [
            (DimCoord([i], long_name="merge_coord1"), None),
            (DimCoord([j], long_name="merge_coord2"), None),
            (aux_coord, None),
        ]
        c = iris.cube.Cube(
            data,
            dim_coords_and_dims=dim_coords_and_dims,
            aux_coords_and_dims=aux_coords_and_dims,
        )
        cubelist.append(c)

cube = cubelist.merge_cube()
print(cube)
print(cube.coord("aux_coord"))

# aux_coord is nan, merged cube contains aux_coord as a 2-dimensional auxiliary coord
aux_coord = iris.coords.AuxCoord([np.nan], long_name="aux_coord")
cubelist = iris.cube.CubeList()
for i in range(2):
    for j in range(2):
        aux_coords_and_dims = [
            (DimCoord([i], long_name="merge_coord1"), None),
            (DimCoord([j], long_name="merge_coord2"), None),
            (aux_coord, None),
        ]
        c = iris.cube.Cube(
            data,
            dim_coords_and_dims=dim_coords_and_dims,
            aux_coords_and_dims=aux_coords_and_dims,
        )
        cubelist.append(c)

cube = cubelist.merge_cube()
print(cube)
print(cube.coord("aux_coord"))

[trexfeathers]

We should consider duplicating the solution in #3283.

Is Iris' array_equal() still appropriate in the places where it is being used?

[stephenworsley]

While the problem is similar, I don't think it is due to a use of array_equal() in this case. It seems as though the problem here is due to comparisons between Cells defined here:

iris/lib/iris/coords.py

Lines 1349 to 1373 in 5eeb2c0

	def __eq__(self, other):
	"""
	Compares Cell equality depending on the type of the object to be
	compared.
	"""
	if isinstance(other, (int, float, np.number)) or hasattr(
	other, "timetuple"
	):
	if self.bound is not None:
	return self.contains_point(other)
	else:
	return self.point == other
	elif isinstance(other, Cell):
	return (self.point == other.point) and (
	self.bound == other.bound or self.bound == other.bound[::-1]
	)
	elif (
	isinstance(other, str)
	and self.bound is None
	and isinstance(self.point, str)
	):
	return self.point == other
	else:
	return NotImplemented

If a Cell containing a NaN was equal to another Cell containing a NaN, I think this would solve our problem. It ought to be possible to do this without changing Cell equality behaviour, though that might involve a bit more of a rewrite of the merge code.

[stephenworsley]

Due to changes in Python 3.10, the approach I took to fix this no longer seems viable. Until we come up with a better solution, #4701 has been reverted so this issue is reopened.