Seven stages, one direction, and exactly one loop. The loop matters: results return through the bus so the orchestrator advances the DAG on an event rather than holding a thread while a task runs.
The Outbox Relay sits between the database and the queue on purpose. Without it the orchestrator would have to write state and send a message as two operations, and there is no ordering of those two that is safe — view 13 walks through why.
Observe is a stage, not a sidebar. The correlation identifiers are minted at admission and carried through every message property, which is what makes view 25's execution waterfall possible at all.
Why these services
Service Bus over Event Hubs or Storage Queues: this is competing-consumer work distribution needing per-message locks, lock renewal for long tasks, scheduled enqueue for backoff, sessions for ordered completion and a native dead-letter queue. Event Hubs is a partitioned log with none of those.
Cosmos DB over Azure SQL for execution state: partition-scoped transactional batches, a change feed that makes the outbox and every projection possible without polling, elastic RU for a spiky workload, and a TTL that expires events without a purge job.
AKS over Container Apps or Durable Functions: the requirement explicitly asks to demonstrate pod disruption, node failure, HPA/KEDA behaviour and rolling deployments. Durable Functions would also make the runtime the orchestrator, which is precisely the coupling view 03 refuses.
Deliberate omissions
Failure paths are absent so the happy path stays readable. Views 16, 19, 20 and 31 carry them.
Backpressure, tenancy and security are not on this page; views 21, 29 and 27 own them.
The Retry Service and Trigger Service exist but are folded into Decide here. View 05 lists every deployable unit.
