An order workflow reserves inventory, authorizes payment, and arranges shipment across separate services. How would you keep the business operation consistent when a later step fails?

I would identify compensable steps, a point after which reversal is impossible or undesirable, and retryable steps that can drive the workflow forward. Irreversible side effects, such as a final external settlement, belong as late as the business allows. Timeouts need durable timers rather than one in-memory delay. I would monitor age by saga state, retry and compensation rates, stuck instances, and end-to-end completion time, with an operator path for cases that require human judgment. If all data still belongs in one service and one transaction, I would keep it there; a saga is justified by a real ownership boundary, not by the appeal of the pattern.