Merging into existing dataset and saving to netcdf throws KeyError

What happened:
I am trying to write a program which adds data to an existing Dataset saved as a netcdf file. Saving a Dataset to a netcdf file, then loading it back and merging it with another Datset works fine. However, when trying to save this combined Dataset to the same filename the program throws KeyError. If another filename is used or if there is no overlap between the dimensions everything works.

What you expected to happen:
I expected the Dataset to overwrite the existing file.

Minimal Complete Verifiable Example:

# Using xarray version 0.16.0
from xarray import DataArray, merge, open_dataset
from numpy import zeros

# First DataArray
da_a = DataArray(zeros((1,1)),coords=[('x',[1]),('a',[2])])
ds_a = da_a.to_dataset(name="var_from_da_a")

# Second DataArray
da_b = DataArray(zeros((1,1)),coords=[('x',[1]),('b',[3])])
ds_b = da_b.to_dataset(name="var_from_da_b")

# This works
ds = merge([ds_a,ds_b])
ds.to_netcdf('works.nc')

# This fails, throws KeyError: 'var_from_da_a'
ds_a.to_netcdf('fails.nc')
with open_dataset('fails.nc') as handle:
    ds_alt = merge([handle,ds_b])
ds_alt.to_netcdf('fails.nc')

Anything else we need to know?:
Copying the handle or manually trying to close the old file (although that should be handled by the "with" statement) does not help.

Also it's my first time opening an issue for xarray so I'm sorry if it's not appropriate and in general I'm really impressed with xarray!

Environment:

Run the supplied minimal example with xarray 0.16.0, then trace the interaction between open_dataset, merge, and to_netcdf when the output filename is reused. Add regression coverage for overlapping dimensions and verify that writing the merged Dataset to the existing filename no longer raises KeyError.