Product Context
ShopNow is a large e-commerce marketplace. Design the recommendation system for the home feed and product-detail-page modules, with a specific focus on handling cold start for both new users and newly listed items.
Scale
| Signal | Value |
|---|
| DAU | 35M |
| Peak recommendation QPS | 180K |
| Active catalog | 120M products |
| New items/day | 2.5M |
| Users with <5 historical events | 28% of DAU |
| End-to-end p99 latency budget | 150ms |
Task
- Clarify the product goal and define success metrics for recommendations under cold-start conditions.
- Propose an end-to-end architecture covering retrieval, ranking, and re-ranking for both warm and cold users/items.
- Design the data and feature pipeline, including what can be computed batch vs near-real-time.
- Choose models for each stage and explain how they address new-user and new-item cold start.
- Define offline and online evaluation, including segment-level analysis for cold-start cohorts.
- Identify major failure modes such as feature drift, training-serving skew, popularity collapse, and stale item metadata.
Constraints
- New products must become eligible for recommendation within 10 minutes of listing.
- Logged-in users may have rich history, but guest users often have only session context, geo, device, and referrer.
- Serving cost must stay below $0.001 per recommendation request.
- Some item attributes from sellers are noisy or missing at listing time.
- The system must support policy filters: out-of-stock items, blocked sellers, and regional compliance restrictions.
Your design should explicitly discuss how to avoid over-relying on popularity for cold start, how to balance exploration vs exploitation, and how to keep training and serving features consistent.
