Divisions by zero in cube arithmetic give inconsistent results

[schlunma]

🐛 Bug Report

The results of zero divisions when performing cube arithmetic depend on cube shapes and whether if they have proper coordinates or not. Sometimes a proper zero-division warning is raised (with an output of nan), sometimes no warning is raised (with an output of -- = masked value). See following code example for more details.

I guess this is somehow related iris' advanced broadcasting using coordinates.

How To Reproduce

Steps to reproduce the behaviour:

import iris.analysis
from iris.cube import Cube
from iris.coords import DimCoord


# Cubes with proper coordinates
x_coord = DimCoord([0], var_name='x')
y_coord = DimCoord([1, 2], var_name='y')
cube = Cube([[0.0, 0.0]], dim_coords_and_dims=[(x_coord, 0), (y_coord, 1)])

# The following raises a zero-division warning
res = cube / cube
print(res.data)  # [[nan, nan]]

# The following raises a zero-division warning
res = cube / cube.collapsed(['x', 'y'], iris.analysis.MEAN)
print(res.data)  # [[nan, nan]]

# The following does NOT raise a zero-division warning
res = cube / cube.collapsed('x', iris.analysis.MEAN)
print(res.data)  # [[--, --]]

# The following does NOT raise a zero-division warning
res = cube / cube.collapsed('y', iris.analysis.MEAN)
print(res.data)  # [[--, --]]


# Cubes without proper coordinates
# These cases ALWAYS raise a zero-division warning

res = Cube([[0.0, 0.0]]) / Cube([[0.0, 0.0]])
print(res.data)  # [[nan, nan]]

res = Cube([[0.0, 0.0]]) / Cube([0.0, 0.0])
print(res.data)  # [[nan, nan]]

res = Cube([[0.0, 0.0]]) / Cube([0.0])
print(res.data)  # [[nan, nan]]

Expected behaviour

All of these cases above should return the same output. I personally have a slight preference towards returning nan values, as this allows us to properly distinguish between a prior mask and the zero divisions.

Environment

• OS & Version: Red Hat Enterprise Linux 8.6 (Ootpa)
• Iris Version: 3.7.0

[valeriupredoi]

this could be sitting a lot deeper in the stack, in netCDF4python layer, there are all manners of surprising side effects when _FillValue is not defined eg Unidata/netcdf4-python#748

[rcomer]

Is this because collapsed returns a cube with a masked array? #5314

[schlunma]

Is this because collapsed returns a cube with a masked array? #5314

Yes, that's the reason! Changing the data to a regular array after collapsed fixes this.

Thanks, I think this can be closed then 👍

[valeriupredoi]

cheers guys! Happy it's not in deeper stacks 😁