Nwm client IndexError: invalid index to scalar variable. about hydrotools

[aaraney]

From the stack trace, it appears that the metadata field, scaling_factor, for the streamflow variables in one of the NWM's channel_rt output files is not being deserialized as a collection (list, etc.) and instead is just a scalar variable (int, float, etc.). This may have been caused by a downstream change to a dependency (xarray, h5netcdf).

[aaraney]

I was able to resolve this issue by removing the index in the scale_factor object.

line 274 python/nwm_client/src/hydrotools/nwm_client/gcp.py

            # Extract scale factor
            scale_factor = ds['streamflow'].scale_factor[0]

            # fixed with
            scale_factor = ds['streamflow'].scale_factor

I am assuming that the metadata layout of NWM channel route link files is pretty static over time as we've not seen this issue before. I assume this is a deserialization issue propagating from, if I had to guess, xarray.

It might be best if we push a hot fix that guards and type checks the scale_factor field while we track down and figure out what is causing this and determine a long term solution.

[aaraney]

Found the issue. It is propagatingh5netcdf. Today they pushed 0.14.0 which introduced the following per their change log.

Return items from 0-dim and one-element 1-dim array attributes. Return multi-element attributes as lists. Return string attributes as Python strings decoded from their respective encoding (utf-8, ascii). By Kai Mühlbauer.

I verified that rolling the version back to 0.13.0 resolved this issue.

[aaraney]

Now as to how we should proceed. I know previously I said:

It might be best if we push a hot fix that guards and type checks the scale_factor field while we track down and figure out what is causing this and determine a long term solution.

In this case, I think it makes sense to just type check ds.streamflow.scale_factor and handle the case where a scalar is returned. I dont want to force others to comply with a version pinning of h5netcdf. Thoughts @jarq6c?

proposed solution

streamflow = ds['streamflow']

# h5netcdf <= 0.13.0 always deserializes numeric attributes to numpy arrays.
# even if there will only be one item in the array.
if isinstance(streamflow.scale_factor, np.ndarray):
  scale_factor = streamflow.scale_factor[0]

# h5netcdf > 0.13.0 deserializes numeric attributes to numpy arrays if there is more than scalar in the attribute.
# otherwise, a  scalar numpy value is returned
else:
  scale_factor = streamflow.scale_factor

[jarq6c]

If the source attribute was a single scalar all along and was only returned in a list because of some conceit of h5netcdf, I'm inclined to just drop the index and leave it at that. Is there a good reason to continue supporting h5netcdf <= 0.13.0?

[aaraney]

After talking with @jarq6c offline, we came to a solution (please correct me where necessary @jarq6c). Given that h5netcdf==0.14.0 was released on 2022-02-25, we will pin the current version of nwm_client (5.0.1) to h5netcdf <= 0.13.0 and release the software as a post release to 5.0.1. Subsequently, nwm_client==5.0.2 will be released and pin h5netcdf >= 0.14.0. 5.0.2 will include a patch that resolves complies with h5netcdf >= 0.14.0.