[lbrandy]

"zero overhead abstractions" is a principle of the language, not the library. The STL containers are clearly designed to be on the "safer" side than the performant side. You are right that, perhaps, we've extracted the wrong lesson from the STL but your version ("iterators everywhere, including for your own algorithms and container types") is still definitely the right message.

As an example, STL unordered-maps have fairly strict requirements around iterator invalidation and that results in indirection that do affect performance. It's pretty easy to make a "faster" unordered_map that tosses that requirement (as many do).