Katkat
Back

Design personalized restaurant search and recommendations

DoorDashMachine Learning Engineer · Technical Screen
MediumMachine Learning EngineerTechnical ScreenSystem Design

Last updated

This question tests the ability to design an end-to-end personalized restaurant recommendation system, covering feature engineering, retrieval-ranking architecture, and infrastructure trade-offs.

Scenario

Design a personalized restaurant recommendation service for a food delivery platform. When a user opens the app, they should see a list of restaurants tailored to their current context, location, and past preferences.

Requirements

  • Functional: Support real-time ranking based on user context (location, time, past orders). Support search queries that combine semantic intent with location constraints.
  • Non-functional: Latency must be under 200ms for the recommendation feed. Handle high concurrency during peak dining hours. System must be scalable to millions of users.

Your Design Should Cover:

  1. Data Model: What entities (User, Restaurant, Order, Context) are stored and how are they retrieved?
  2. Architecture: Describe the retrieval-ranking pipeline. How do you handle candidate generation versus the final ranking model?
  3. ML Strategy: What features are used? How do you update the model online vs. offline? Where can an LLM/semantic search add value compared to traditional collaborative filtering?
  4. Trade-offs: Discuss the balance between latency and model complexity (e.g., how to handle re-ranking at scale).
  5. Evaluation: How do you measure the success of the recommendations (online vs. offline metrics)?

Hints

Hint 1

Consider a two-stage architecture where candidate generation retrieves hundreds of restaurants from a candidate pool and a ranking model scores the top few.

Hint 2

Think about using vector databases for semantic search to handle natural language queries like 'spicy Thai food near me'.

Hint 3

Address the cold-start problem for new restaurants or users by discussing content-based features versus interaction-based collaborative filtering.

Follow-up questions

  • How would you ensure the feed remains diverse and avoids 'filter bubbles' where users only see one type of cuisine?
  • If you deploy a new model, how would you implement a canary rollout or A/B test to measure its impact without breaking the user experience?
  • How would the system design change if we needed to incorporate real-time inventory updates or temporary restaurant closures due to high order volume?