This seems especially appealing in the awkward middle: too serious for in-memory queues, not big enough to justify Kafka-shaped machinery.