That's not a bad point, and perhaps I should test creation times too. (I wanted to ignore that s.t. the test isolated a single thing: searching / wasn't mixing two things together.) Both hashsets and vectors are O(n) in setup, and I'd mostly expect their real-world performance to be very similar: hash tables tend to use contiguous arrays for storage and vectors do by definition, both tend to overallocate (hash tables to prevent collisions, vectors to amortize appends to O(1)), so I'd expect the performance in the real-world to be similar. Vectors tend to insert in-order, though, whereas the insertions in hashsets would bounce around the table, so that might make it more interesting.