Design Product Analytics for Billion-Event Daily Ingestion

Design an in-house product analytics system for mobile and web applications. Provide an SDK and ingestion API, enrich and store events, support fast minute-level count dashboards, and support more complex warehouse queries. ### Requirements - About one billion events arrive per day; traffic spikes need not be modeled. - A dashboard can request counts per minute for an event type, such as installs over the last two hours. - Any such count query covering data from the past week must finish in under500 ms for99% of requests. - Enrichment can add or update analysis fields, such as reverse-geographic information or context tags. - A warehouse supports queries such as page-load counts grouped by U.S. state over the last three months. - Handle out-of-order arrivals and duplicate events. ### Clarifying Questions Define event-time versus ingestion-time reporting, acceptable freshness, deduplication identity and retention, dimensions available on the fast path, and late-event correction policy. The fast dashboard requirement is minute counts, not arbitrary low-latency warehouse SQL. ### What a Strong Answer Covers Capacity estimates, durable ingestion, SDK retry behavior, separate raw and serving stores, minute aggregates, enrichment versioning, late-event updates, deduplication, and evidence for the stated latency target. ### Follow-up Questions - How do you avoid counting an SDK retry twice? - What happens when a late event changes a minute bucket already shown on a dashboard? - Why should enrichment failures not necessarily block basic event counts?