Upstream test for NumPy arrays to inlined Kerchunk references fails

[maxrjones]

Full error available in https://github.com/zarr-developers/VirtualiZarr/actions/runs/13959203618/job/39077253124

[TomNicholas]

Oh cool that's 2 fewer tests that are my fault in #477 haha

[maxrjones]

The relevant update is in pydata/xarray#9498, which changes the return dtype from np_arr = CFDatetimeCoder().encode(var.copy(), name=var_name).values float32 to float64 in our tests. Since the original dtype was datetime64[ns], the new behavior seems like an improvement to me at first glance. I'll look into what would need to be changed in Kerchunk to match Xarray's new behavior.

[maxrjones]

Well Kerchunk's also not wrong to keep it as float32 since that's how it's stored on disk 😕 Not sure the correct solution since there's no np.datetime32 to work with when decoding times. I think we're probably popping the necessary dtype encoding somewhere, still looking 👀

from netCDF4 import Dataset
import xarray as xr

file = "test_numpy_arrays_to_inlined.nc"
ds = xr.tutorial.open_dataset("air_temperature")

ds.to_netcdf(file, format="NETCDF4")

nc_file = Dataset(file, 'r')


for var_name, var in nc_file.variables.items():
    print(f"Variable: {var_name}, Data type: {var.datatype}")

Variable: lat, Data type: float32
Variable: air, Data type: int16
Variable: lon, Data type: float32
Variable: time, Data type: float32