Design an e-commerce recommendation system
Overview
This question evaluates expertise in large-scale machine learning system design, covering competencies in recommendation algorithms, candidate generation and ranking, feature engineering and stores, low-latency serving, and operational reliability; it is categorized in the ML system design domain and requires both conceptual understanding and practical application. It is commonly asked to assess the ability to balance personalization and business metrics (CTR, CVR, revenue), address scalability and latency targets, and manage data quality, cold-start, exploration–exploitation, A/B testing, bias/fairness, and monitoring trade-offs in production.
Design an Amazon-Scale E‑Commerce Product Recommendation System
Context
You are designing a large-scale recommendation system that powers multiple user touchpoints in an e‑commerce platform. The system must handle high traffic and a very large catalog, deliver low-latency personalized recommendations, and be robust to data issues and model drift.
Requirements
-
Primary Use Cases (surfaces)
- Home feed (personalized recommendations)
- Product detail page (PDP: similar/related items, cross‑sell, up‑sell)
- Cart/checkout (complements, bundles, substitutions)
- Email/notifications (batch recommendations)
-
Success Metrics
- Core: CTR, add-to-cart rate, conversion rate (CVR), revenue per mille (RPM), expected GMV, margin-adjusted revenue
- Experience: diversity, novelty, freshness/recency, coverage (long tail), personalization lift
- Reliability: latency p50/p95/p99, error rate
- Guardrails: bounce rate, returns/cancellations, customer satisfaction proxies (NPS/CSAT), seller/category fairness
-
Architecture
- Candidate generation and ranking services
- Key features: user behavior, item content/metadata, graph/co‑view/co‑buy signals, context
- Model choices: MF/BPR, deep two‑tower retrieval, sequence models, gradient‑boosted trees, DLRM/MLP re‑rankers
- Feature store (offline + online, point‑in‑time correctness)
-
Pipelines and Serving
- Offline training pipelines and batch feature generation
- Near‑real‑time updates to features and models where appropriate
- Online serving, caching strategy, latency/SLA targets, and scale estimates
-
Special Topics
- Cold start for users and items
- Long‑tail coverage and catalog exploration
- Exploration–exploitation (e.g., bandits)
- A/B testing design and guardrails
-
Risk, Compliance, and Ops
- Data quality, feedback loops, debiasing
- Bias/fairness and privacy
- Abuse/fraud prevention
- Monitoring, alerting, and rollback plans
Solution
Login required