Choose Retrieval vs Ranking Stack

Product Context

ShopNow is a large e-commerce marketplace with a personalized home feed and search result page. The team is debating whether to introduce a two-tower retrieval model or continue scaling a more traditional learning-to-rank stack built on hand-engineered candidate generation and a pointwise ranker.

Scale

Task

Design the end-to-end recommendation/ranking system and explain when you would choose a two-tower model vs a traditional ranking approach for this product.

1. Clarify the product objective and define success metrics for home feed and search-like surfaces.
2. Propose a multi-stage architecture, including candidate generation, ranking, and optional re-ranking, and explain where a two-tower model fits.
3. Compare at least two design options: (a) traditional retrieval + ranker and (b) learned two-tower retrieval + downstream ranker. State the tradeoffs in recall, latency, freshness, cost, and maintainability.
4. Design the training and serving stack, including online vs batch features, feature store usage, index refresh, and fallback behavior.
5. Define offline and online evaluation, including how you would detect if retrieval improves recall but hurts business metrics downstream.
6. Identify key failure modes such as feature drift, training-serving skew, cold start, and stale embeddings, with mitigations.

Constraints

• New items must become eligible for recommendation within 10 minutes of catalog ingestion.
• 35% of traffic comes from anonymous or weak-history users.
• Serving cost must stay under $0.001 per request at peak.
• Some seller and pricing features are restricted and can only be used in aggregated form for compliance reasons.
• The system must support both personalized browsing and high-intent query sessions without maintaining separate infrastructure unless clearly justified.