issue_comments

CC @fujiisoup who implemented much of this. I will also take a look at your doc when I have the chance.

I do think that handling duplicate matches with indexing is an important use-case. This comes up with nearest neighbor matching as well -- it would be useful to be able to return the full set of matches within a given distance, not just the nearest match.

I wonder if it would be more productive to consider a new indexing API for one -> many matches. sel/loc is already quite complex.

Please take a look at xarray's detailed indexing rules: http://xarray.pydata.org/en/stable/indexing.html#indexing-rules

I will ignore the dimension names for now as I don't have much experience with xarray yet.

I think this is the crux of the problem. Put another way: why should the result of indexing be a 1x2 array instead of a 2x1 array? Currently (with the exception of indexing by a scalar with an index with duplicates), xarray determines the shape/dimensions resulting from indexing from the shape/dimensions of the indexers not the array being indexed.

Let me give a more concrete example of the issue for multi-dimensional indexing: python da_unique = xr.DataArray([0, 1], dims=['x'], coords={'x': ['a', 'b']}) da_nonunique = xr.DataArray([0, 1], dims=['x'], coords={'x': ['a', 'a']}) indexer = xr.DataArray([['a']], dims=['y', 'z']) With a unique index, notice how the result takes on the dimensions of the indexer: ```

da_unique.loc[indexer] <xarray.DataArray (y: 1, z: 1)> array([[0]]) Coordinates: x (y, z) object 'a' Dimensions without coordinates: y, z `` What would you propose for the result ofda_nonunique.loc[indexer]`?

Thanks for the report!

This was actually a somewhat intentional omission in xarray, but if would not be particularly difficult to add in this feature if we want it. At the very least, we should note this deviation somewhere in the docs.

There are two potentially problematic aspects to the pandas behavior: 1. It means that you cannot count on indexing a dataframe with its own index to return something equivalent to the original dataframe, e.g., consider df.loc[['a', 'a']] in your example, which returns a dataframe with 4 rows. 2. More generally, it means you can't count on indexing a dataframe with an array to return an object of the same size as the indexer. This is particularly problematic for xarray, because we support vectorized indexing with multi-dimensional indexers. I don't know how we could define a multi-dimensional equivalent of this -- what shape should the result have if you indexed with a multi-dimensional array instead, e.g., da.loc[{"dim_0": xr.DataArray([['a']]}]? With multiple dimensions involved, it's not clear where the extra introduced dimensions should go.

Now that you bring this up, I wonder how the existing supporting for indexing like da.loc[{"dim_0": "a"}] would work if there are other multi-dimensional indexers. I don't know if we have test coverage for this...