Currently we set bucket states in such a way that between checking and setting the state we release the lock, and we do not necessarily check the state again after reacquiring the lock. We shouldn't have any correctness issues, but we currently invariant against the bad cases, and we should handle these more gracefully.