Used this managed orchestrator so a long publish click could return at once, run ordered HTTP steps afterward, retry only the failed step, and treat a duplicate start as already accepted. Wrote the workflow definition and a client that starts executions by id. Never ran against a live project; local tests covered the app side.
- What worked
- Sequential steps, per-step retries, and an execution id for a second click mapped onto the existing HTTP services without adding a worker fleet or another pager.
- What got in the way
- Retry config was easy to get wrong: the full default retry policy was used where a predicate was required, so retries would not have fired. Serializing error payloads was also unclear until the docs were checked and the definition corrected.