Should InvertedIndices call to_indices on all elements in an array of skipped indices?

[bkamins]

We have the following super bad behavior:

julia> [1, 2, 3, 4][Not(Not([1.5]))]
1-element Vector{Int64}:
 2

julia> [1, 2, 3, 4][Not(Not([1.5]))]
1-element Vector{Int64}:
 0

julia> [1, 2, 3, 4][Not(Not([1.5]))]
1-element Vector{Int64}:
 4294967297

which is a consequence of the fact that:

to_index(I::AbstractArray) = I

in Base.

@mbauman - do you think we should fix it in InvertedIndices.jl or this should be fixed in Base?

CC @nalimilan

[bkamins]

Another issue with Not:

julia> x = [1, 2, 3]
3-element Vector{Int64}:
 1
 2
 3

julia> x[Not(Integer[true, 2])]
1-element Vector{Int64}:
 3

and this should error

[mbauman]

Fundamentally, the challenge here is that to_indices assumes that an array of N indices will generate N scalar getindex calls (or some cartesian product thereof). And if they're not the canonical index types, those scalar getindex calls will happily call to_indices once again.

But what can we do if we're skipping those elements? More fundamentally, how do we prove that a skipped index would otherwise be a valid index — and which one is it? Should we pre-emptively map to_indices over all arrays of skipped indices? Maybe only if they'd change? Or should we try to do it lazily (that's #41; it's pretty ugly)?

I am slightly curious how DataFrames is avoiding the troubles that #7 displays.

[bkamins]

DataFrames.jl provides a custom implementation of Not for column selection.