RFC: Concatenating iterator

[mschauer]

An iterator that concatenates the elements of its argument, see #16622 . For example

julia> collect(Base.concat((i:i+2 for i in 1:4)))
3×4 Array{Int64,2}:
 1  2  3  4
 2  3  4  5
 3  4  5  6

It produces the elements in the same order as flatten. To determine the shape of the children it peaks at the first element of the argument. This implementation works for iterators which have the HasShape trait. Iterators with HasLength trait can be incorporated in a next step if the proposed design is sound.

[tkelman]

throw(ArgumentError("argument to Concat must contain at least one element")) would be more consistent with other iterator constructor error conditions

[pabloferz]

You could write it like this collect(concat(i:i+10 for i in 1:3)) wouldn't you?

[JeffBezanson]

Thinking about this some more, I'm not sure it adds much value over Flatten since it can only concatenate along trailing dimensions. It's also not good practice to call start before the user of the iterator, which may point to an inherent difficulty with n-d concatenation via iterators. Instead we might want a function f(itr, dim) that concatenates the elements of itr along a specified dimension. I'm not yet sure what to call this though.

[mschauer]

The one thing what I found is nice about having a iterator implementing this (at the cost of peeking at the first element) is that it combines well with other iterators. Concatenation along a different dimension becomes

julia> collect(concat(zip([1:3,1:3]...)))
2×3 Array{Int64,2}:
 1  2  3
 1  2  3

and concatenation two levels deep is concat(concat(...)) etc

[mschauer]

Maybe it is not an atypical situation that information about length becomes known as soon as the iterator got started and not before -- and that is all what is needed in collect, something like a length(itr, state)-function.

[Keno]

Is there a utility of having this as a shaped iterator as opposed to just reduce(hcat, (i:i+2 for i in 1:4))? In any case, I'm gonna close this since it's written against the old iterator protocol. If this is to be revived, it can be a new PR (with the new iteration protocol) or go in a package.