issue_comments: 36216552

Notes: 1. DO NOT MERGE -- some of the units tests are currently failing (see below). 2. I decided to rebase to use some of @ebrevdo's recent changes. The discussion on the previous commit can be found here: https://github.com/akleeman/xray/commit/d700eeb4da836cf300e92af6d9dcf17427d67e50

Failing tests:

I added an "encoded_dtype" attribute to keep of the original dtype of variables loaded from a netCDF file. Unfortunately, this means most of the round-trip tests currently fail, because no variables have "encoded_dtype" attributes until they are loaded from netCDF files. I think we will need to make some ugly trade-off to get round tripping working both directions, but I'm not yet sure what the best option is. We could: 1. Ignore some specific attributes (like "encoded_dtype") when checking XArray equality. 2. Save these encoding details on XArrays outside of the attributes dict. 3. Add special logic for the round-trip tests to ignore these attributes.

FWIW, I think it's OK if we don't preserve data types in the round-trip process, as long as the data itself is equivalent. I'm not entirely opposed to trying, but in general it is very hard to guarantee that serialized/unserialized data is exactly equivalent. There is somewhat of a conflict between preserving the original data (netCDF-like) and representing the data in a sane format in-memory (which should not be exactly like a netCDF). IMHO, we should focus on the later.