Fixes for compat='no_conflicts' and open_mfdataset

[shoyer]

• xarray.merge now uses compat='no_conflicts' by default.
• compat='no_conflicts' preserves attributes.
• Variable.no_conflicts broadcasts arguments, making it a strictly more
  relaxed check than broadcast_equals.
• Allow concat_dim=None in open_mfdataset and auto_combine to disable
  trying to infer a dimension for concatenating.
• Add compat as an argument to open_mfdataset to allow more control over
  merging.

CC @jcmgray

[jcmgray]

Should be 'no_conflicts' and 'broadcast_equals' here I think.

[shoyer]

oops, thanks

[jcmgray]

This all looks great to me. Might the docstring for auto_combine need to be updated? Currently it seems to suggest that sorting, alignment and the addition of new data variables are not supported, whereas I think all these can occur with compat='no_conflicts'.

[shoyer]

Might the docstring for auto_combine need to be updated? Currently it seems to suggest that sorting, alignment and the addition of new data variables are not supported, whereas I think all these can occur with compat='no_conflicts'.

Yep, with compat='no_conflicts' and recent align updates to concat, we are most of the way there on alignment.

It still can fail if there are different variables on datasets across the concatenated dimension, but that will be fixed when we get concat to handle different variables (#508). At that point, we might want to switch the logic from the current "concatenate by variable group, then merge" to "merge by unique coordinates, then concatenate".