The interview report described two system design rounds: a webhook delivery platform and a vehicle metrics aggregation system. ### Constraints & Assumptions - For webhooks, assume multi-tenant external delivery over HTTPS with at-least-once semantics. - For telemetry, assume many connected vehicles send measurements every few seconds. - Discuss APIs, storage, queueing, retries, security, observability, scaling, and fault tolerance. - Make delivery and aggregation tradeoffs explicit. ### Clarifying Questions to Ask - For webhooks, what event volume, tenant count, latency target, payload size, and ordering requirements exist? - Can customers replay missed webhook events? - For telemetry, how many vehicles and events per second are expected? - What retention and query latency targets are required? - Are alerts real-time, batch, or both? ### Part 1 - Design Webhook Delivery How would you design a multi-tenant webhook platform? #### What This Part Should Cover - APIs for endpoints, subscriptions, secrets, logs, and replay. - Event bus, subscription lookup, fanout, delivery queue, workers, retry queue, dead-letter queue, and observability. - At-least-once delivery, idempotency, signing, tenant isolation, and failing endpoint protection. ### Part 2 - Design Vehicle Metrics Aggregation How would you design a telemetry aggregation platform for connected vehicles? #### What This Part Should Cover - Ingestion gateway, stream platform, stream processing, raw storage, hot aggregate store, dashboard serving, alerting, and batch reprocessing. - Handling duplicates, late events, out-of-order events, missing data, retention, and downsampling. ### What a Strong Answer Covers - Defines delivery semantics and failure behavior. - Separates critical delivery or ingestion paths from analytics. - Chooses durable queues and replayable storage. - Discusses observability and tenant or fleet isolation. ### Follow-up Questions - Why is exactly-once webhook delivery unrealistic? - How would you rate-limit one bad customer endpoint? - How would you handle out-of-order telemetry? - What storage would you use for recent dashboards versus long-term raw data? - How would you replay historical telemetry after changing aggregation logic?