NetCDF attributes like `long_name` and `units` lost on `.mean()`

[j08lue]

When reading in a variable from netCDF, the standard attributes like long_name, standard_name, and units are being propagated, but apparently lost when calling .load() .mean() on the DataArray.

Couldn't these CF-Highly Recommended Variable Attributes be kept during this operation?

(What to do with them afterwards, e.g. upon merge, is a different question, unresolved also in the pandas community.)

EDIT: the problem actually occurs when calling .mean() (not .load(), as originally posted).

[shoyer]

Hmm. This is definitely a bug -- load should preserve all metadata.

[shoyer]

Could you post an example dataset/code for which this occurs? I'm struggling to reproduce this.

[j08lue]

Sorry for the confusion! The loss of attributes actually occurs when applying .mean() (rather than .load()).

See this notebook (same in nbviewer) for an example with some opendap-hosted data.

[shoyer]

Ah. So this is intentional. There is an optional parameter that lets you control this -- try .mean(keep_attrs=True).

The basic problem is that it's ambiguous how to handle attributes like units after doing computation. I don't want to inspect attributes and choose some to preserve and others to remove, so we have a choice of either preserving all attributes in an operation or removing all of them.

Obviously, for some aggregations (e.g., sum or var) it doesn't make sense to preserve attributes (which commonly include units). I suppose we could make an exception for aggregations like mean/median/std, but it's also weird to have some aggregations that preserve attributes and others that don't.

[j08lue]

That makes sense. Great that there is an option to keep_attrs. Closing this issue.