When is FastAPI `BackgroundTasks` enough, and when do you need a durable worker?

I define the delivery contract explicitly: at-most-once best effort, at-least-once with idempotent effects, or a reconciled external operation. Queue presence alone does not create exactly-once behavior. Backpressure protects dependencies, retry schedules distinguish transient from permanent failure, and metrics cover age, attempts, failure class, and completion. During shutdown, in-process tasks may be cut short, while durable workers need lease or visibility-timeout semantics. The architectural choice follows business consequence and recovery needs, not the number of lines required to configure a queue.