System Design: Real-Time Top‑K Purchased Items Over Rolling Windows

Design a real-time service that ingests purchase events and continuously outputs the top‑K most purchased items over configurable rolling windows (for example: last 5 minutes and last 24 hours).

Input Event

Each purchase event contains:

• customer_id
• item_id
• timestamp (event time)

Assume K and window sizes are configurable at runtime.

Requirements

Specify and justify:

1. Data model for events and results.
2. Ingestion pipeline and partitioning.
3. Aggregation/windowing strategy (e.g., event-time keyed windows, panes).
4. Algorithms/data structures to maintain top‑K at scale (e.g., heaps, Space‑Saving, Count–Min Sketch + heap), including complexity and memory trade‑offs.
5. Handling late, duplicate, and out‑of‑order events (watermarks, allowed lateness, deduplication).
6. State management, TTL/retention, and backpressure control.
7. Fault tolerance and exactly‑once processing semantics end‑to‑end.
8. Horizontal scalability and hot‑key mitigation.
9. APIs for querying the current top‑K for given windows (and update cadence/latency).

Make minimal, explicit assumptions as needed (e.g., target throughput, latency SLOs), and propose a design that works for both a short window (5m) and a long window (24h).