If you're talking about the discussion of head and snoc at the beginning of the article, laziness just moved the costs around - head becomes O(n) (from O(1)), but snoc becomes O(1) (from O(n)), because you only pay for the snoc costs when inspecting the queue with 'head'.

It can be stunning the first few times you see this sort of thing, but Haskell run time performance is not that hard to reason about it - is simply different from strict languages. As an experienced Haskeller I could see their point at the definition of the data structure and operations. In fact there is an entire data structure library built around this idea of O(1) snoc/appends, the dlist: https://hackage.haskell.org/package/dlist-0.7.1.2