When displaying a MaskedArray, I'm told the data, the mask, and the fill value. Of course, data and mask are very important. But what is the practical significance of the fill value? I can even change it, but why would I want to do that — isn't the fill value just an implementation detail with no practical impact?
In other words: does the fill_value have any impact on any code not directly addressing fill_value?
Looking at the Masked_Array class code, I see:
methods for setting and getting fill_value
filled() method, which returns a copy with the masked values replaced by the fill_value. This is the 'direct' use of it.
methods that call filled as part of their calculation.
masked.all() fills with True and then does the ordinary array all.
masked.any() fills with False.
masked.nonzero() does:
return narray(self.filled(0), copy=False).nonzero()
trace and sum also fill with 0, but prod fills with 1.
argsort (and other methods like argmin) uses:
d = self.filled(fill_value).view(ndarray)
Those methods take a fill_value parameter, or use the self.fill_value. For methods like this the user potentially has strong preferences as to how the masked values are used in sorting or taking the minimum/maximum.
So fill_value is essential to efficient array calculation. Some methods require a special value, others can use whatever the user wants.