Design personalized restaurant search and recommendations
Quick Overview
This question evaluates a candidate's expertise in designing scalable, personalized restaurant search and recommendation systems, including system architecture, recommendation model design, LLM integration points, API and service boundaries, data storage and online/offline pipeline trade-offs, and operational concerns like latency, reliability, and evaluation metrics. It is commonly asked in system design interviews for machine learning engineering roles to probe practical, application-level skills in recommendation systems and real-time search, emphasizing applied architectural reasoning and engineering trade-offs rather than purely conceptual theory.
Scenario
You are designing a DoorDash-like personalized restaurant recommendation system.
A user types a free-text query (e.g., “spicy ramen under $20”, “healthy vegetarian lunch”, “best burgers near me”). The system should return a ranked list of restaurants (optionally with dishes) personalized to the user.
What to design
- End-to-end architecture for query → ranked restaurants.
- How you would incorporate LLMs into the system (where they help, where they shouldn’t be in the critical path).
- Key APIs / services , main data stores , and online vs offline components.
- Ranking approach (candidate generation + re-ranking), personalization signals, and dealing with constraints (delivery radius, store hours, availability, price, dietary needs).
- Latency & reliability targets and strategies (caching, fallbacks, degradation).
- Evaluation : offline metrics, online A/B metrics, and guardrails.
Assumptions to clarify
- You have restaurant metadata (location, cuisine, menu items, prices), user history, and real-time availability.
- Traffic is large (e.g., millions of daily users). Typical p95 latency target for search/recs is ~200–500ms.