combine_first of Datasets changes dtype of variable present only in one Dataset

What happened: I was combining two Datasets using combine_first and to my surprise the dtype of one of the DataArrays in the merged Dataset was changed (from bool to float64).

What you expected to happen: No change in dtype.

Minimal Complete Verifiable Example:

>>> import xarray as xr
>>> ds = xr.Dataset(coords={'abc': list('abc')})
>>> ds['x'] = ('abc', [1., 2., 3.])
>>> ds['y'] = ('abc', [-1., -2., -3.])
>>> ds['t'] = ('abc', [True, False, True])
>>> ds
<xarray.Dataset>
Dimensions:  (abc: 3)
Coordinates:
  * abc      (abc) <U1 'a' 'b' 'c'
Data variables:
    x        (abc) float64 1.0 2.0 3.0
    y        (abc) float64 -1.0 -2.0 -3.0
    t        (abc) bool True False True
>>> xy4b = ds[['x', 'y']].sel(abc=~ds.t) * 10
>>> xy4b.combine_first(ds)
Out[14]: 
<xarray.Dataset>
Dimensions:  (abc: 3)
Coordinates:
  * abc      (abc) object 'a' 'b' 'c'
Data variables:
    x        (abc) float64 1.0 20.0 3.0
    y        (abc) float64 -1.0 -20.0 -3.0
    t        (abc) float64 1.0 0.0 1.0

Anything else we need to know?: No.

Environment:

No source file or test is named; reproduce the supplied Dataset.combine_first example first, then trace the combine_first implementation. Done means the result keeps variable t as bool rather than float64, with a regression check for the example.