test/sort.jl broken on current Julia master

[martinholters]

Since JuliaLang/julia@a4e3cd2, typed_vcat (called from vcat) explicitly converts its arguments to Arrays. This breaks isequal(sort(da), [DataArray(sort(dropna(da))); DataArray(T, nna)]) since the [...; ...] tries to put NAtypes into an Array{Float64,1}.

[tkelman]

If the best solution to this is to merge revert JuliaLang/julia#17079 for 0.5, please let us know quickly.

[ararslan]

Someone else surely would know better than I would (maybe @andreasnoack? he knows everything) but I'm not sure what we could do in the short term on our end to accomodate that change.

[andreasnoack]

The reasoning behind JuliaLang/julia#17079 was that full(a) would always be convert(Array, a) but that is clearly not true so the solution should either be to revert it modify it but given how close we are to 0.5, reverting might be the right choice.

[martinholters]

Actually, the fact that, for some da::DataArray{T}, isa(da, AbstractArray{T}) and hence eltype(da) == T, but potentially !isa(da[idx], T) looks like it is breaking a lot of assumptions. Funny it doesn't trigger more errors.

That said, the simplest solution for this particular issue might be adding specialized methods to vcat (and probably the other cat methods, too).

Or, even simpler, disallow cat'ing DataArrays and rewrite the offending tests 😄

EDIT: I had completely missed JuliaLang/julia#17564...

[nalimilan]

This is annoying. I think this shows that DataArrays are fundamentally broken as they don't respect the element type they declare. OTOH breaking DataArrays in Julia 0.5 would be a serious issue since the replacement isn't completely ready yet. So either we're able to override all cat methods for DataArrays, or we should keep JuliaLang/julia#17079 reverted until 0.6.

[andreasnoack]

JuliaLang/julia@a4e3cd2 Has been reverted so this is not an issue anymore. I'm wondering if the element type of a DataArray should be Union{T,NAType}. That seems right to me and would work with JuliaLang/julia@a4e3cd2