With not so important and small cosmetic change (If Stefan change word “probably” => “maybe”) his claim seems to me to be correct. (in short version: Julia is pretty good 
)
But I also think that warning in “Performance tips” and “Notable difference from python” about generators (and maybe also about strings) are still useful.
I got about a 50x iteration time speed-up (comparing directly) replacing the dict with vectors and using @views on a better sized test set (~10000 items, with a max three links per node)
49x for xloop = 1, 85x for xloop = 10
Julia doesn’t seem to allocate if the set is small enough, even with less than optimal @codewarn_type
add: ~30x against the (cheat-ish) non-allocating 10,000x11 approximation, 0.06 sec to 0.002x sec