handling vectors with `Missing` in type signature (even if there are no missing values)

[kescobo]

julia> using Makie

julia> x = rand(10); y=rand(10);

julia> scene = Scene();

julia> scatter!(scene, x, y, markersize=0.5);

julia> scatter!(scene, x, Union{Missing, Float64}[y...], color=:red, markersize=0.5);

julia> scene

@jkrumbiegel noted on slack:

it turns the yvalues into ascending integers

[jkrumbiegel]

Ok I found the bug. Here we convert our input arrays according to some traits:
https://github.com/JuliaPlots/AbstractPlotting.jl/blob/master/src/conversions.jl#L131

and the trait system treats arrays with number type elements as continuous:
https://github.com/JuliaPlots/AbstractPlotting.jl/blob/master/src/conversions.jl#L105-L114

Which means that a range from 1:n is generated as y-values because the system assumes our Union{Missing, Float64} values are categorical..

[jkrumbiegel]

Don't know how to fix this cleanly, though

[SimonDanisch]

Hm, more fine grained method overloads?

[jkrumbiegel]

there could be another filter step introduced or just an error if there's missing data. what's better?

[SimonDanisch]

I think it'd be best, if element type Union{Missing, Number} would never dispatch to the categorical conversion function ;)

[asinghvi17]

The problem then becomes:

extrema([1, 2, missing, 6])
(missing, missing)

so we'd have the axis limits as missing. We could fix this by implementing a filter, or by dropping any missings in the position vectors?

[kescobo]

julia> extrema([1, 2, missing, 6] |> skipmissing)
(1, 6)

?

[asinghvi17]

Actually we could replace missing with nan, that should be fine

[rapus95]

Well, if you replace missing with NaN you make the step from
Union{Int64, Missing} to Union{Int64, Float64}. Is that any better? I guess the best approach mostly depends on the use case. Usually simply skipping will do it. Sometimes a vertical line which somewhat looks like a "cut" would be a valid option aswell. So I'd recommend to accept some argument which will be coalesced into the array. And in case that's a missing just call skipmissing.

[rapus95]

How about adding the rule:

categorical_trait(T::Type{>: Missing}) = categorical_trait(Base.nonmissingtype(T))

[SimonDanisch]

julia> extrema([1, 2, missing, 6] |> skipmissing)

Or update this definition to work with missing: https://github.com/JuliaPlots/AbstractPlotting.jl/blob/f298fe6d66399639fe04c1fc906d1f289189edd1/src/layouting/data_limits.jl#L42

And then make sure it's used everywhere ;)

[rapus95]

Sure, that's needed either way as I understand it! But the trait would still be resolved wrongly, wouldn't it?

[SimonDanisch]

Yeah, this was just a comment about that particular problem!

[SimonDanisch]

How about adding the rule:

categorical_trait(T::Type{>: Missing}) = categorical_trait(Base.nonmissingtype(T))

Did you try it? Could maybe work?