I don't believe this is a case of nonidempotent callbacks, unless you are referring to Future.set_result(), which by design can't be called twice.  The callbacks are given an inconsistent opportunity to modify the poll set because of indeterminacy in the ioloop.  That being said I understand your reluctance given the amount of turmoil this has but would argue that consistency with tornado is a powerful ally and that a model where any callback using call_soon will be guaranteed the opportunity to modify the poll set is a good thing.