Design real-time top-K products service

Question

Design a real-time service that ingests a continuous stream of purchase events with the schema (customer_id, item_id, timestamp) and continuously computes and serves the top-K most purchased items. The service must support multiple rolling windows at once, for example:

• A short sliding window (e.g. last 5 minutes, or a configurable window T such as the last 1 hour), and
• A long window and daily rolling totals (e.g. last 24 hours, plus per-calendar-day totals with both a running "today" view and finalized prior days).

Work through the full design and address each of the following:

1. Data model for the ingested event and for the materialized top-K result records.
2. Ingestion pipeline and partitioning — durable log, partition key choice, and hot-key mitigation.
3. Windowing and aggregation strategy — sliding windows vs. tumbling micro-buckets with add/subtract, keyed windows, slide cadence, and how daily totals are finalized.
4. Algorithms and data structures for maintaining top-K at scale (e.g. count map + min-heap for exact, or approximate heavy-hitters such as Space-Saving / Misra-Gries and Count-Min Sketch + heap), including the two-stage local-then-global merge pattern.
5. Event-time processing with watermarks: handling late, out-of-order, and duplicate events; allowed-lateness budgets per window.
6. State management and TTL — embedded state store, checkpointing, per-window namespaces, and eviction.
7. Fault tolerance and exactly-once vs. at-least-once semantics.
8. Horizontal scalability — partitioning, key-splitting for skew, autoscaling.
9. Serving and query APIs — how clients fetch the current top-K per window/day, plus optional streaming push.
10. Latency vs. accuracy trade-offs and how you would tune them.