Kai Mühlbauer

@ohsqueezy You might also try `engine="h5netcdf` (`h5py`/`h5netcdf` packages needed). And would it be possible create a small subset of that file via netCDF4 to share?

> To fix this, I think logic in `_choose_float_dtype` should be updated to look at `encoding['dtype']` (if available) instead of `dtype`, in order to understand how the data was originally...

New attempt here: #8713

@ohsqueezy After #7654 going out of focus I've resurrected the changes in #8713. Would you mind taking a look there, maybe test and let us know your thoughts on that?

@fsvenson Thanks for raising this. Could you share the full error traceback please?

Thanks @fsvenson, that's helpful. So if you just want to get back the reference we could do this as suggested. I've no idea about the implications this might have downstream....

I looks like netcdf4-python is just ignoring attributes with type reference. In the past `h5netcdf` added these features which are unsupported in netcdf4/netcdf-c as `invalid_netcdf`. Looking at https://github.com/h5netcdf/h5netcdf/blob/18c86f414cc2915a78cc5f30c25ab5f65bae2139/h5netcdf/core.py#L1127-L1148 `dtype="reference"` is...

@fsvenson Sorry, this fell through the cracks. Are you still in the works with this? Would you be able to create a MCVE producing such file (eg via h5py) or...

@ZedThree This might also be of interest for you.

@ZedThree That's my information too. The question is, why netcdf4-python is using transient types within the dataset instead of using shared committed type?