NA support for Vector{Float}

[juliadata]

NA support for Vector{Float}

[tshort]

I've played with this some more. One area of concern has been display and user confusion. For the most part, that can be minimized. It's a three-line change to base to allow separate display of NaN's and NA's. So, even though there are different internal NA representations, the user can mainly use NA, and everything will be handled transparently.

Another issue I found is with floating-point comparisons. With NA's (or NaN's), a floating point comparison always evaluates to false. Here's an example:

julia> 1.0 .< [NA, 3., NA]
3-element Bool Array:
 false
  true
 false

julia> 1.0 .>= [NA, 3., NA]
3-element Bool Array:
 false
 false
 false

In practice, I don't think it's a big deal, but it is confusing. Pandas has this same issue, and I don't see where they even document it.

A separate floating-point bitstype could be created to solve this, but for comparison operations, it would have to return a type that could contain NA's. R users are used to handling the NA case in comparisons, but here, if one forgets, answers could be wrong.

[garborg]

Closing in favor of recent discussion about NullableArray design.