Katkat
Back

Design a Low-Latency Store Recommender

DoorDashData Scientist · Onsite
HardData ScientistOnsiteMachine Learning

Last updated

This question assesses the ability to architect a multi-stage recommendation pipeline that balances low-latency retrieval with high-accuracy ranking in a geo-aware, high-concurrency food delivery context.

Scenario

DoorDash needs to recommend stores to a user within 50ms of a request. The user is browsing in a specific geohash, and we have millions of stores with dynamic features (e.g., store delivery time, promotional status) and user features (e.g., recent order history, dietary preferences).

Task

Design a scalable system that provides real-time store recommendations. Your response should address the following areas:

  1. Pipeline Architecture: Describe the stages from candidate retrieval to final ranking (e.g., how do you handle millions of stores to produce a top-k list?).
  2. Feature Engineering & Serving: How do you manage real-time updates for features like 'current estimated delivery time' or 'store busy status' while maintaining sub-millisecond lookups?
  3. Geo-Sharding: How does the system handle geospatial constraints? How do you partition the data to ensure the store index is queryable by current user location?
  4. Non-Functional Requirements: Address how you handle high request volume, system cold starts, and model deployment (e.g., A/B testing and canary rollouts).

Constraints

  • Latency budget: < 50ms end-to-end.
  • Traffic: 100k requests/second at peak.
  • Consistency: Store availability must be highly accurate to prevent negative user experience.

Deliverables

  • A high-level system diagram description.
  • A brief justification for your choice of retrieval and ranking models.
  • Strategies for caching vs. computation trade-offs.

Hints

Hint 1

Consider using a two-tower architecture or approximate nearest neighbor (ANN) search for the retrieval phase to narrow the candidate pool efficiently.

Hint 2

Think about the difference between static features (user location history) and transient features (store kitchen capacity) and how to serve them via different data stores (e.g., Key-Value stores vs. Feature Stores).

Hint 3

Evaluate the trade-off between strict consistency and eventual consistency for real-time signals like 'Store Busy' status.

Follow-up questions

  • How would you measure the performance of this system using offline metrics (NDCG, Recall) versus online metrics (Conversion Rate, Mean Time to Order)?
  • If you detect a significant shift in store data distribution, what automated mechanisms would you implement to trigger model retraining or adjustments?
  • How would you modify the ranking layer to ensure diversity in the recommendations (e.g., avoiding only showing fast-food chains)?