Coupang ML System Design: Scalable E-commerce Search

You are asked to design a scalable search system for a large e-commerce platform (Coupang-scale). The system must take user text queries, handle millions of products, and return relevant results within strict latency and availability targets. You should cover query processing, candidate retrieval (keyword and semantic/vector), ranking, filtering/facets, autocomplete, and optional personalization.

Core content: explain how queries are parsed, tokenized, and normalized; how spelling correction, synonyms, and intent signals are applied; and how two retrieval paths (inverted-index keyword search and vector/semantic search using ANN) generate candidate sets. Describe a two-stage ranking pipeline where a fast lightweight model or heuristic filters candidates, and a richer ML ranking model (learning-to-rank / gradient boosting / neural ranker) reorders results using relevance, popularity, inventory, and user signals.

Flow / stages you should discuss:

• Query ingestion & front-end: rate limiting, geo-routing, and autocomplete service
• Candidate retrieval: inverted indices, sharded ANN/vector indexes, hybrid fusion
• Feature store & ranking: online features, coarse-to-fine re-ranker, A/B testing hooks
• Filtering & faceting: attribute indexes and fast post-filtering
• Serving, caching & observability: edge caches, cold-start handling, monitoring and alarms

Skill signals: demonstrate knowledge of information retrieval (inverted index), semantic/vector search (embeddings + ANN), ML ranking (LTR), capacity planning for thousands QPS, latency budgeting (95th percentile <200ms), high availability patterns, sharding/replication, and trade-offs between recall, precision, and freshness.