issue_comments

Closed by #2876

Updated title. I'm happy with the explicit approach. New design prototype: ``` wallet, restaurant_bill = xarray.align(wallet, restaurant_bill, join='outer', fill_value=0) print(wallet - restaurant_bill)

<xarray.DataArray (currency: 2)> array([50, 40]) Coordinates: * currency (currency) object 'EUR' 'USD' ```

This behavior is consistent with the default behavior on pandas, which always does an outer join for arithmetic: In [6]: wallet.to_series() - restaurant_bill.to_series() currency EUR NaN USD 40.0 dtype: float64 So I don't think we want to change this in an inconsistent way in xarray.

I do agree that we should have support for an explicit fill value in alignment/reindexing and arithmetic. For consistency with pandas (and elsewhere in xarray), let's call it fill_value.

• change __add__,__sub__ etc. to invoke xarray.align(fillvalue=0)
• change __mul__, __truediv__ etc. to invoke xarray.align(fillvalue=1)

I can see the logic in using an identity value instead of NaN as a default in arithmetic. One peril of this approach is that it isn't always evident what the right identity is. In fact, according to NumPy: ``` In [16]: import numpy as np

In [17]: print(np.add.identity) 0

In [18]: print(np.multiply.identity) 1

In [19]: print(np.subtract.identity) None

In [20]: print(np.divide.identity) None ```

Let me give a couple other examples of why I don't think we should use an identity of some sort as the default: - Suppose we are comparing a model to observations. model - observations gives the residuals -- unless the variables are mis-aligned, in which case it would give a residual equal to one of the variables. - Suppose we want to do a weighted average. (observations * weights).sum() gives a sensible answer -- but not if a missing weight defaults to 1 (all the weights put together are supposed to sum to 1!).