eBay ML System Design: Post-Checkout Recommendations

You are asked to design a post-checkout recommendation system for an e-commerce platform (eBay) that must surface exactly six product recommendations immediately after a user completes checkout. The primary goal is to maximize revenue: a recommendation only pays when a user clicks and purchases the recommended item, so you must optimize for both click-through and conversion rates while meeting strict latency and scale requirements.

Start by describing a two-stage architecture: a fast candidate retrieval stage (ANN, category filters, popularity & recency heuristics) that reduces millions of items to a few hundred candidates, followed by a latency-aware ranking model that scores and returns exactly six items. Explain feature engineering choices: user embeddings from purchase history, item embeddings (ID embeddings + TF-IDF for titles/descriptions), contextual features (time of day, device), price and margin features, and user-item interaction features (recent co-purchases, time-decayed counts). Discuss handling high-cardinality categorical features and sparse vectors to remain cost-effective (hashing, compressed embeddings, quantization).

Cover model training and labels: label clicks and purchases separately (multi-task or cascade), use pairwise or listwise ranking loss, and evaluate offline with NDCG/AUC and online with CTR, conversion rate, and revenue-per-session via A/B tests. Address low-latency serving (100 ms budget) with precomputed embeddings, feature store lookups, efficient ANN indexes, caching, and fallback rules. Finally, explain data ingestion (transaction logs, catalog updates, user profiles), labeling pipelines, retraining cadence, observability (metrics, drift detection), and how your design supports maintainability and incremental updates.