Make AxisKeys play nicely with AcceleratedArrays about axiskeys.jl HOT 1 OPEN

I agree it would be nice to make this all work. Trying to recover what's actually going on first, this works:

a = AcceleratedArrays.Interval(sd, ed)
A(in(a), :) # in(...) isa Function, so this goes to:
findall(in(a), dates) # isa StepRange, could be UnitRange

but without the in it gets treated as a scalar, which fails:

a |> typeof |> supertype # , isn't anything recognised
AxisKeys.findindex(a, axiskeys(A,1)) # error, from this:
findfirst(isequal(a), axiskeys(A,1))

For IntervalSets.Interval there's a special findindex method to avoid this:
https://github.com/mcabbott/AxisKeys.jl/blob/master/src/selectors.jl#L6
It doesn't call findall(in(s), r), because that would be an error:

findall(in(IntervalSets.Interval(sd, ed)), dates.parent)
findall(in(IntervalSets.Interval(3,5.5)), 0:10) # same error without Dates

and instead falls back to giving a vector, as you observe. The issue to fix that is JuliaMath/IntervalSets.jl#52 (which should become a PR).

But it remains to figure out what to do with AcceleratedArrays.Interval, without depending on that package. Why does this differ anyway, it's some issue about isless vs <? But it could still subtype AbstractInterval, that would be nice because IntervalSets.jl is a small package for precisely this purpose.

Is there another way to catch this? We could send z::Any to finall(in(z), axiskeys(A,1)) instead of findfirst(isequal(z), axiskeys(A,1)), i.e. treat objects whose type doesn't match eltype of keys as collections, not individual values, but that doesn't seem ideal. For example push!(KeyedArray(rand(3), [:a, :b, :c]), 0.44) has eltype Any.

from axiskeys.jl.