Too many attributes converted to string

[aulemahal]

In save_to_zarr and save_to_netcdf, we coerce attributes of "funky" types into strings before writing to file. However, our "funky" check is not exhaustive enough.

See:

xscen/xscen/io.py

Lines 333 to 338 in 134c31b

	if not (
	isinstance(attrs[k], (str, float, int, np.ndarray))
	or isinstance(attrs[k], (tuple, list))
	and isinstance(attrs[k][0], (str, float, int))
	):
	attrs[k] = str(attrs[k])

An example is found here:

import xarray as xr

ds = xr.open_dataset('/you/know/where/caq/invariants/areacella_caq_fx.nc')
type(ds.rotated_pole.attrs['grid_north_pole_longitude'])

Which give : numpy.float32 , something that the line above does not catch. And thus, floating point coordinates of the rotated pole are converted to string when written to file, which breaks it.

I suspect something changed in the recent xarray, because I otherwise don't understand how we wouldn't have caught that earlier.

[RondeauG]

That doesn't quite seem to be the problem. I just tested it and isinstance(attrs[k], (str, float, int, np.ndarray)) does catch it correctly and keeps it as a numpy.float32. ncview crashes nonetheless.

I also tried to force either str or float (non-Numpy), but without success in ncview.

[aulemahal]

Darn.
On my end isinstance(np.float32(0), (str, float, int, np.ndarray) is False and the ncview error explicitly said: NetCDF: Attempt to convert between text & numbers. I am guessing there are 2 issues at play here, depending on the package versions maybe.

I'm on:
numpy 1.23.5
xarray 2022.12.0
xscen 0.4.10-beta

[RondeauG]

I'm on the latest xscen (0.4.12), but numpy 1.22.4 and xarray 2022.6.0

[aulemahal]

Woups, I'm sorry, I had not noticed that with my code, both methods seem to work.

With to_netcdf, the attributes are indeed in the file as numbers.
With xs.save_to_netcdf, attributes are saved as strings.

However, in both cases, ncview opens and I have not seen different behaviours.

EDIT:
ncview understands the save_to_netcdf version as rlat is latitude and rlon is longitude. If I plot coastlines over the data (see Opts), I actually see Africa instead of North America.