You already know: failure models, timeouts and idempotency (Distributed systems fundamentals); streaming windows, event time and watermarks (D11 ML pipelines); journals and effectively-once side effects (D9 Advanced agent engineering).
Why this matters for Atlas. Messaging decouples producers from consumers: the grader publishes "quiz answered" once and moves on; progress, features, analytics and alerts each consume it at their own pace, and a consumer that's down catches up later instead of failing the grader. Almost every AI platform runs on this: feature pipelines, training-data collection, inference request queues, agent event triggers, model monitoring. The hard part is semantics: ordering, what happens on failures (lost or duplicated messages), and how to make effects happen exactly once when delivery can't be. This lesson builds a partitioned log like Kafka's, consumer groups and offsets, and the patterns (idempotent consumers, transactional outbox, dead-letter queues) that make it correct.
1. Queues versus logs
The question. What's the difference between a message queue and an event log, and which does Atlas need?
Start from nothing. Two designs: