Product Context
ShopSphere is a large e-commerce marketplace. You are designing the cross-sell recommendation system that shows "Frequently bought together" and "You may also need" suggestions on product detail pages, cart pages, and post-add-to-cart flows.
Scale
| Signal | Value |
|---|
| DAU | 35M |
| Peak recommendation QPS | 120K |
| Active product catalog | 25M SKUs |
| Daily orders | 4.5M |
| New or updated SKUs/day | 1.2M |
| Per-request latency budget (p99) | 120ms |
Cross-sell suggestions should improve attach rate and revenue per session without hurting the primary purchase funnel. Recommendations must work for both logged-in and logged-out users, and should adapt to rapidly changing inventory, price, and promotions.
Deliverables
- Clarify the product objective, request surfaces, and success metrics for cross-sell recommendations.
- Design an end-to-end ML system, including candidate generation, ranking, and any re-ranking or business-rule layer.
- Define the offline and online data pipelines, labels, features, and training cadence.
- Specify the online serving architecture, including latency budget, caching, fallbacks, and capacity planning.
- Propose an evaluation and experimentation plan, plus monitoring for drift, training-serving skew, and system health.
- Identify major failure modes such as cold start, out-of-stock items, stale features, and popularity bias.
Constraints
- Must exclude out-of-stock, restricted, or incompatible items before final ranking.
- Inventory and price changes can happen within minutes; stale recommendations are costly.
- 40% of traffic is logged-out, so personalization is limited on many requests.
- Serving cost should stay below $0.001 per recommendation request.
- Some categories (e.g. healthcare, age-restricted items) require policy filtering and auditability.
- The system should degrade gracefully to non-personalized recommendations if user features or models are unavailable.
