issue_comments

Sorry for taking so long here.

I'm still not entirely comfortable with this change -- it adds some additional complexity to the Dataset constructor that seems unnecessary. There's no inherit reason why calling a constructor on an object of the same type needs to work. If anything, I would like to deprecate/remove this behavior on DataArray (e.g., in favor of using a more explicit call to DataArray.copy(data=new_data)).

To answer your question though, we hit this when inheriting from a Dataset: the object is initialized with a dataset, which is super-ed up to the Dataset constructor

If I understand this correctly, you've written something like: python class MyDataset(xarray.Dataset): def __init__(self, dataset): super().__init__(dataset)

Is there any reason why you couldn't just write: python class MyDataset(xarray.Dataset): def __init__(self, dataset): super().__init__(dataset.data_vars, dataset.coords, dataset.attrs)

I think we should do something here, rather than change behavior by coincidence.

I don't think it's quite a coincidence here -- we are changing the behavior of Dataset.__iter__/Dataset.keys(), and this is a logical consequence of interpreting a Dataset as a mapping of its data variables :).

Now I'm having second thoughts. It does make me a little nervous to add special cases that distinguish between xr.Dataset(ds, coords=some_coords) and xr.Dataset(ds.data_vars, coords=some_coords).

Maybe my proposed solution above is too complex, and it would be better simply to default the coords and attrs argument to those arguments from a passed Dataset object. This is basically what we do for DataArray.

That said, I do think we should try to resist the notion that every class constructor must also work for coercion to that class. This tends to result in a slightly confused interface. A separate as_dataset() function (even in user code) could give exactly the desired result in only be a few lines of code.

An alternative approach - more friendly but less explicit - is for any other argument (coords, attrs) supplied to overwrite the Dataset's

I think this would be slightly preferred, and also more consistent with the model of a Dataset as only iterating over data_vars. This suggests that at the very least, Dataset(ds) and Dataset(ds.data_vars) should be consistent after we finish the changes described in https://github.com/pydata/xarray/issues/884. Note that ds.data_vars will pull in coordinates associated with the DataArray objects.

To be more precise, I would probably make: ds2 = Dataset(ds1, coords, attrs) equivalent to ds2 = ds1.copy() ds2.coords.update(coords) ds2.attrs.update(attrs)