xr.DataArray.sum() converts string objects into unicode

What happened:

When summing over all axes of a DataArray with strings of dtype object, the result is a one-size unicode DataArray.

What you expected to happen:

I expected the summation would preserve the dtype, meaning the one-size DataArray would be of dtype object

Minimal Complete Verifiable Example:

ds = xr.DataArray('a', [range(3), range(3)]).astype(object)
ds.sum()

Output

  <xarray.DataArray ()>
    array('aaaaaaaaa', dtype='<U9')

On the other hand, when summing over one dimension only, the dtype is preserved

ds.sum('dim_0')

Output:

   <xarray.DataArray (dim_1: 3)>
    array(['aaa', 'aaa', 'aaa'], dtype=object)
    Coordinates:
      * dim_1    (dim_1) int64 0 1 2

Anything else we need to know?:

The problem becomes relevant as soon as dask is used in the workflow. Dask expects the aggregated DataArray to be of dtype object which will likely lead to errors in the operations to follow.

Probably the behavior comes from creating a new DataArray after the reduction with np.sum() (which itself leads results in a pure python string).

Environment:

Start by running the provided DataArray.sum() example and compare the all-axis reduction with the one-dimension reduction. Trace the DataArray.sum() path through np.sum(), then add a regression test showing that an all-axis sum of object strings retains object dtype and remains compatible with dask.