Feed Ranking

Source: 30-System-Design/Framework & Systems

Prompt

• Rank posts for a social feed given a user, session context, and candidate set.

Requirements

• Functional: relevance, freshness, diversity, creator fairness.
• Non-functional: low latency, robustness to spam, transparency.
• Constraints: real-time updates, abuse resistance, content safety.

Success Metrics

• Offline: AUC, NDCG, calibration.
• Online: session length, return rate, hide/report rate.
• Guardrails: complaints, block rates, policy violations.

Data

• Sources: user interactions, graph signals, content features.
• Labeling: clicks, dwell time, follows, hides.
• Position bias: log impressions and position for debiasing.

Modeling

• Candidate generation: graph-based retrieval, collaborative filtering.
• Ranking: deep ranker with engagement prediction.
• Re-ranking: diversity, freshness, creator caps.
• Session context: intent, recency, time-of-day features.

Serving

• Retrieval: cache + ANN + graph store.
• Latency budget: multi-stage ranking with timeouts.
• Real-time: session features and recency updates.
• Fallback: trending or follow-based feed if personalization fails.

Evaluation & Monitoring

• Offline eval: temporal splits, position bias correction.
• Online eval: A/B tests with safety guardrails.
• Drift/abuse: feedback loops, adversarial content.
• Quality: monitor creator distribution and long-tail coverage.

Risks & Tradeoffs

• Engagement vs well-being.
• Freshness vs personalization.
• Diversity vs relevance in short sessions.

Notes