Remove "no non-missing" warning for character min and max?

[mattdowle]

When looking at gsumm.c for recent PR, this warning struck me as unnecessary :

> DT
        a     b
   <char> <int>
1:   <NA>     1
2:   <NA>     2
> DT[, min(a, na.rm=TRUE), by=b]
       b     V1
   <int> <char>
1:     1   <NA>
2:     2   <NA>
Warning message:
In gmin(a, na.rm = TRUE) :
  No non-missing values found in at least one group. Returning 'NA' for such groups to be consistent with base

Can we just remove the warning? I can imagine that the warning is more annoying when it does occur, than being appreciated. A secondary motivation is that there is an extra loop through the result just to generate that warning, so that loop could be removed.

[jangorecki]

IMO definitely! If it is consistent to base R then we don't have to warn. This is what user should expect.

[mattdowle]

The PR didn't turn out the way I was planning: the warnings are gone, but it no longer returns Inf. I realize you agreed on the basis that the bahavior was the same as base so why warn. But then I realized the warning itself is generated in base too (that Inf is being returned). So with-warning is the part that is consistent with base too.
I just think the new behavior (NA) makes more sense. I'm sure everyone is surprised by the Inf and type change to numeric when they first experience it. Time to remove that user surprise. NA is the behavior I want myself when I'm doing data science.

But it's still open for discussion. It's merged in dev with a long news item so folk can test dev. All the way up until release I'm happy to reconsider.