#undef in SentinelArray

[bkamins]

This is a follow up to JuliaData/Tables.jl#186 to keep track of the issue.

@quinnj - does it make sense to define a custom isassigned for SentinelArray? I think it is not a problem and should simplify things to add it, but you know the design much better.

[quinnj]

Do you have a specific idea for what isassigned would mean here? Like, are you thinking it would check for #undef, even if it's used as the sentinel, like for SentinelArray(["hey", "there"])? The isassigned fallback works already, but we could define Base.isassigned(A::SentinelArray, i) = isassigned(parent, i) to "pass-through" to the more efficient Array implementation under the hood.

[bkamins]

We probably misunderstood the SentinelArray design there. So in practice if parent has a possibility of #undef (i.e. it is not a bits type) it will always get turned to missing and be used as a sentinel - right?

If so (i.e. it is not possible to have #undef in SentinelArray) isassigned could just do what Base.checkbounds returns I think and then isassigned would be fast.

[quinnj]

You can still have #undef in SentinelArray, but you would have to use your own sentinel manually, i.e. by default, reference types will use #undef as their sentinel. So, for example, you can still do SentinelArray(Vector{String}(undef, 3), "custom_sentinel") and get #undef elements.

[bkamins]

I guess we could check:

1. if isbitstype || sentinel is #undef - it is enough to check range else fall back to the default implementation (it should be rare in practice)