issue_comments: 1385842347

background

(Moving the convo out of a comment for visibility)

For reference, the code we would like is python array = as_indexable(self.array) array = array[self.key] array = array.get_duck_array()

problem

So far I have found two fail cases

1. Wrapping a backend array instance

This method is removing LazilyIndexedArray. When we do that, we sometimes have another ExplicitlyIndexed array (_ElementwiseFunctionArray) and sometimes a BackendArray. We then apply array[self.key] which returns a ExplicitlyIndexed array for the former and forces a load from disk for the latter.

One way to solve this would be to return another ExplicitlyIndexed array from the getitem on the BackendArrays. This is currently forcing a load from disk:

I tried this by making __getitem__ on a BackendArrayWrapper return a ExplicitlyIndexed array (commit). This breaks all existing backends, and is a little complicated to add to the Zarr and Scipy backends.

2. Wrapping an IndexingAdapter instance

The other case is when as_indexable returns an IndexingAdapter instance. Again, these return duck arrays from __getitem__, so we can't call get_duck_array

https://github.com/pydata/xarray/blob/522ee2210499cfa434a0ca0825790949a8e08ff0/xarray/core/indexing.py#L671-L690

how to proceed?

@shoyer I'm not sure on how to proceed. 1. For BackendArrays we could wrap them in another lazy indexing class when the Variable is created so we don't break our backend entrypoint contract (that __getitem__ can return a duck array). We might need something like this for TensorStore anyway. 2. I don't know what to do about the indexing adapters.