Skip identical indexes with non-unique values in align?

[shoyer]

Currently, when objects with with non-unique (duplicated) values in one of their indexes are passed to align, an error surfaces from pandas:
InvalidIndexError: Reindexing only valid with uniquely valued Index objects

We could certainly give a more informative error here (see this complaint on StackOverflow), but a bigger issue is that this probably isn't strictly necessary. Instead, we could skip indexes for alignment if they are already equal. This is slightly less principled (a non-unique index may indicate something has gone wrong), but certainly more convenient and more inline with how pandas works (e.g., it even allows arithmetic between objects with non-unique indexes, which I believe does not work currently in xarray).

Currently, we do this as a special case when merging arrays and exactly one has labels (see _align_for_merge in #950). But we could probably do this in general, either by default or with a flag to enable it (or turn it off). This would then propagate to every xarray operation that uses align under the covers.

[max-sixty]

Another option is to fully disallow non-unique indexes. Not sure how big a use case this is, so this might be a non-starter.

But two wondrous features of xarray:

• Smaller & more defined surface than pandas, so not forced to have all these work arounds
• Coords that aren't dimensions, so labels are possible in place of indexes

[shoyer]

I think it's important to preserve the ability to work with non-unique indexes insofar as it is necessary for cleaning your data. That's why we support non-unique indexes in merge, which is internally used anytime you assign to a Dataset. I'm less sure about the use cases for arithmetic and the like.

[shoyer]

I just checked on xarray v0.7.2 and the example from the StackOverflow post worked. I think we should treat this as a regression -- we were already doing this prior to v0.8.