allow specifying a fill value per variable

[keewis]

While working on #4163 I noticed that the fill value parameter for align (but maybe also reindex, concat, merge and combine_*?) will be used for all variables (except dimension coordinates) which is obviously not ideal when working with quantities. Would it make sense to optionally allow fill_value to be a dict which maps a fill value to a variable name?

Consider this:

In [2]: a = xr.Dataset( 
   ...:     data_vars={"a": ("x", [12, 14, 13, 10, 8])}, 
   ...:     coords={"x": [-2, -1, 0, 1, 2], "u": ("x", [-20, -10, 0, 10, 20])},  
   ...: ) 
   ...: b = xr.Dataset( 
   ...:     data_vars={"b": ("x", [7, 9, 3])}, 
   ...:     coords={"x": [0, 3, 4], "u": ("x", [0, 30, 40])}, 
   ...:  
   ...: ) 
   ...:  
   ...: xr.align(a, b, join="outer", fill_value=-50)
Out[2]: 
(<xarray.Dataset>
 Dimensions:  (x: 7)
 Coordinates:
   * x        (x) int64 -2 -1 0 1 2 3 4
     u        (x) int64 -20 -10 0 10 20 -50 -50
 Data variables:
     a        (x) int64 12 14 13 10 8 -50 -50,
 <xarray.Dataset>
 Dimensions:  (x: 7)
 Coordinates:
   * x        (x) int64 -2 -1 0 1 2 3 4
     u        (x) int64 -50 -50 0 -50 -50 30 40
 Data variables:
     b        (x) int64 -50 -50 7 -50 -50 9 3)

I'd like to be able to do something like this instead:

In [3]: xr.align(a, b, join="outer", fill_value={"a": -30, "b": -40, "u": -50})
Out[3]: 
(<xarray.Dataset>
 Dimensions:  (x: 7)
 Coordinates:
   * x        (x) int64 -2 -1 0 1 2 3 4
     u        (x) int64 -20 -10 0 10 20 -50 -50
 Data variables:
     a        (x) int64 12 14 13 10 8 -30 -30,
 <xarray.Dataset>
 Dimensions:  (x: 7)
 Coordinates:
   * x        (x) int64 -2 -1 0 1 2 3 4
     u        (x) int64 -40 -40 0 -40 -40 30 40
 Data variables:
     b        (x) int64 -50 -50 7 -50 -50 9 3)

I could get there by passing the default (dtypes.NA) and then using fillna, but that only seems to work with data variables so coordinates would need to pass through a reset_coords / set_coords cycle. Also, with this the dtype is changed to float.

[shoyer]

Would it make sense to optionally allow fill_value to be a dict which maps a fill value to a variable name?

Yes, this sounds like a welcome improvement!