Design Real-Time Recommendation System — Student Lab

Offline system design lab: build a real-time home feed recommender architecture with latency/freshness constraints.

Prompt

Design a real-time home feed recommender.

Constraints:

• ●p95 end-to-end latency: 200ms
• ●freshness: new content visible within minutes
• ●candidate pool: millions
• ●strict safety/policy filtering
• ●avoid feedback loops and spammy optimization

0 — Requirements and Metrics

Fill:

• ●primary objective metric
• ●4–8 guardrails
• ●latency SLOs (p50/p95/p99)
• ●online vs offline evaluation plan

1 — Data, Labels, and Feature Freshness

Specify:

• ●events to log (impression/click/dwell/hide/report)
• ●label definitions and delayed labels
• ●feature tiers: realtime / nearline / batch
• ●point-in-time correctness strategy

2 — Candidate Generation (Retrieval)

Design a candidate generation strategy that can return ~1k candidates in <50ms. Include:

• ●ANN index or graph-based retrieval
• ●freshness injection
• ●safety/policy filtering stage

3 — Ranking and Serving Architecture

Design ranking that fits within the remaining latency budget. Include:

• ●model family
• ●feature fetch strategy (online store/cache)
• ●caching strategy (candidates, features, scores)
• ●fallbacks

4 — Experimentation, Monitoring, Rollback

5 — 5-minute summary