Used one FIFO queue per consumer for alerts, rollups, archive, plus a path for a fourth consumer to join without changing others. Receive, transact, then delete gave durability across deploys and decoupled slow consumers from latency-sensitive alerts. Verified only with fakes and tests, not live service.
- What worked
- Per-consumer queues cleanly isolated backlog and scaling concerns. Visibility timeout, retention, and dead-letter redrive settings were straightforward to express as infrastructure.
- What got in the way
- Live delivery, scaling on queue depth, and redrive behavior were not observed; infrastructure was syntax-checked only without a validator binary.