issue_comments

this should be ready for review

@AyrtonB, I took the liberty of pushing the changes I had in mind to your branch, using a adapted version of your docstring. The only thing that should be missing is to figure out if it's possible to reduce the number of computes to 2 instead of n_columns + 1.

@pydata/xarray, any opinion on the API design?

I don't think there is a lot left to decide: we want to keep the conversion logic in from_dask_dataframe and maybe helper functions, and I think we should mirror the pandas integration as close as possible (which means we need a Dataset.from_dask_dataframe and a DataArray.from_dask_series class method).

The only thing I think is left to figure out is how to best compute the chunk sizes with as few computations of dask (as defined by raise_if_dask_computes) as possible.

cc @dcherian

there's a few things to fix in pycompat for this to work: first of all, import dask.dataframe before accessing dask.dataframe.core.DataFrame. We should also move the assignment to dask_dataframe_type to its own try / except block since it's possible to have dask.array but not dask.dataframe installed. And the reason for the ImportError you got is that we need a value for dask_dataframe_type if there was a ImportError. I'm thinking of something like this: ```python try: import dask.dataframe

dask_dataframe_type = (dask.dataframe.core.DataFrame,)

except ImportError: dask_dataframe_type = () ```

sorry, it is indeed called dask_array_compat. Looking closer, you probably won't be able to use that. Instead, I'd advise to do a local import (for an example see Dataset.to_dask_dataframe). For the change in variable.py I would use the same pattern as for pycompat.dask_array_type, so if dask.dataframe is not available dask_dataframe_type should be ().

you can just decorate tests that require dask with requires_dask and they will be skipped automatically if dask is not installed

Edit: actually, you seem to import dask in some modules, which is not what we want. We usually use either the dask_compat module or pycompat.dask_array_type to work around that.