Engineering handoff · Northwind Ingestion
Take ownership without rediscovering the failure boundaries
This handoff gives the incoming owner the current runtime path, safe controls, in-flight decisions, and the first week of proof.
Current status in one minute
- The production path still uses one queue,
ingest, with no failover. - RFC 014 is proposed and not shipped; ADR 009 is accepted for its partition key.
- The 2026-07-30 incident recovered through consumer scaling and remains the representative failure.
- Release candidate
sha-8132c3is go-with-waivers for the current architecture.
Prove the local service before changing it
15 minRun the committed test path
bashmake ingest-testSuccess: gateway contract and consumer tests pass without production credentials.
Start the local gateway
bashmake ingest-runSuccess: logs contain
listeningon port 8443.Send one synthetic event
bashcurl -k https://127.0.0.1:8443/events \ -d '{"id":"evt_handoff","type":"handoff.ping"}'Success: response is
202and repeatsevt_handoff.
Where truth lives
ownersBehavior
openapi/events.yaml
Producer contract. It wins over the HTML companion.
Runtime
Deployment manifests
Replica counts, queue URI, health checks, and environment names.
Operations
Ingestion / lag
Lag, queue depth, request p99, shed count, and runbook link.
Tripwires and in-flight work
do not breakDo not claim the split path is live. The dispatcher, two queues, and INGEST_SHED response remain proposed.
- Never change the event-ID idempotency rule without an API review.
- Never delete a queue as rollback or incident mitigation.
- Reliability owns approval of shed-versus-block behavior.
- The catalog backfill is the explicit opportunity cost if RFC 014 is funded.
First-week checklist
proof- Run the local proof above.
- Shadow one Platform on-call review and locate the escalation policy.
- Read RFC 014, ADR 009, the API companion, and the latest postmortem in that order.
- Reconcile the dashboard queue name with the deployment manifest.
- Write down any access gap; do not place credentials in this document.
Handoff is complete when: the incoming owner can explain the current and proposed paths, run the local proof, identify the safe incident control, and name the two unresolved RFC decisions.