t would be most useful to this particular user if the error message could somehow identify the file(s) responsible for the exception.

+1. You could make this change to open_dataset and it will be raised with open_mfdataset too. Attempting to read a bad netCDF file is a common source of trouble. So an error saying something like

Reading file XXX failed. The file is possibly corrupted, or the file path is wrong.

would be quite helpful!

I would find it very useful if I could pass to open_mfdataset some kind of argument that would make it ignore invalid files altogether (ignore_invalid=False comes to mind).

This I'm not sure about because a user wouldn't know if they were missing some data in the middle...

My vote is to have both, a warning, and an option to fill missing data with NaNs. My use case:

I have an archive of 15 years of monthly forecasts. For one month one of the ensemble members is missing. I am converting binary format to zarr. The code is:

ds = xr.open_mfdataset(
    paths,
    engine=BinaryBackend,
    dtype=np.float32,
    combine="nested",
    concat_dim=(ensmem_ix, fcsttime_ix, reftime_ix),
    parallel=False,
).rename_vars(foo="sic")

Currently, my only option is to remove the remaining ensemble member data files before processing. Since I have to use a custom backend (based on https://github.com/aurghs/xarray-backend-tutorial/tree/main), I tried to add code to return array filled with nans when np.fromfile() fails. That, however, is not enough, the missing file is also accessed in _chunk_ds() in xarray/backends/api.py, to create a token for dask. That could be easily handled by adding a try ... except block.

Contributions welcome!