In my experience the failure rate is far below 5% for vanilla HM algorithm. I've never seem a pathological input in the wild that wasn't contrived specifically to test the typechecker.