Detect missing in Dict constructor

[bkamins]

We have that array constructors correctly detect missing and create a proper Union type of the resulting object:

julia> [1,2,missing, 1.0]
4-element Array{Union{Missing, Float64},1}:
 1.0
 2.0
  missing
 1.0

However, Dict constructor does not follow this pattern and uses Any:

julia> Dict(1=>1, 2=>missing)
Dict{Int64,Any} with 2 entries:
  2 => missing
  1 => 1

julia> Dict(1=>1, missing=>2)
Dict{Any,Int64} with 2 entries:
  missing => 2
  1       => 1

but if we wrap the pairs in a vector it does:

julia> Dict([1=>1, 2=>missing])
Dict{Int64,Union{Missing, Int64}} with 2 entries:
  2 => missing
  1 => 1

julia> Dict([1=>1, missing=>2])
Dict{Union{Missing, Int64},Int64} with 2 entries:
  missing => 2
  1       => 1

My proposal is to make Dict constructors consistent and always create an appropriate Union.

I see the reason in the source code why currently we have this bad behavior, i.e. eltype for <:Tuple uses typejoin and not promote_type (like arrays do), but I could not find any efficient way to fix it.

Of course eltype for tuple should use typejoin, but then maybe typejoin(::Missing,T) should return Union{Missing, T} if !(T>:Missing). I do not know if this was discussed and this is breaking.

CC @nalimilan

[nalimilan]

Yes, that inconsistency is annoying. eltype for tuples should also return Union{T, Missing}. Should be fixed by either #25423 or #24332, depending on which solution we eventually retain.

It could be worth thinking more generally about whether it's OK that Array constructor uses promote but not Dict.

[nalimilan]

Actually, looks like Dict(1=>1, missing=>2) will need more changes, as there are currently specialized methods for when pairs all have the same type, and the fallback is Any. Shouldn't be hard to fix, but since it's a separate issue I won't fix it in the PRs I mentioned above to keep things simple.

[bkamins]

Yes, but I would update constructors after we have a target behavior of typejoin because this will influence the design.

[JeffBezanson]

The == method for dicts could also be updated to handle missing.

[bkamins]

I have a related question (I could not locate it discussed; if it was and there is some conclusions or if there is a better place to discuss it I would be obliged if you let me know). The problem is that broadcast also loses union with Missing, e.g.:

julia> x = [1, missing, 2]
3-element Array{Union{Missing, Int64},1}:
 1
  missing
 2

julia> identity.(x)
3-element Array{Any,1}:
 1
  missing
 2

[bkamins]

Sorry - now I see - it is #25423 - right?

[nalimilan]

Yes, that's it.

[nalimilan]

The == method for dicts could also be updated to handle missing.

See #25661.

[nalimilan]

See #25805 for the Dict constructors.