Complementary item retrieval
Your task
Design the online candidate-retrieval stage for an e-commerce 'complete the look' / related-items service that, given an item a user is viewing, must return ~50 complementary items in under 20ms p99 at 25k QPS over a 400M-item catalog. Complementary (goes-with) is different from similar, a shirt should retrieve pants, not other shirts. Walk through the retrieval architecture, how you serve it within the latency budget, and how you keep results fresh as inventory and prices change every minute.
It covers
- Candidate retrieval
- Two tower
- Ann fanout
- Fallback