Observe the complete workflow
Track latency, failures, accepted outcomes and cost at the workflow boundary. Assign owners to alerts and define a response process. Logging every event without a useful question creates cost and privacy exposure without necessarily improving the service.
Recover without repeating harm
A retry can be safe for reading a document and unsafe for writing a payment instruction. Define idempotency, checkpoints and approval behaviour around consequential actions. Test a timeout, partial execution, stale credential and unavailable dependency before widening access.
Keep the operating model explicit
Managed services can remove some infrastructure work while leaving application, data and outcome responsibilities with the customer. A hybrid or self-managed choice changes the boundary rather than eliminating it. Record who patches, evaluates, monitors, supports and pays for each layer.