I think the point was that it's hard to make something failure-tolerant without being distributed. A locally persisted queue would avoid certain failure modes, though.