Katkat
Back

Design a ranking and recommendation system

MetaSoftware Engineer · Technical Screen
MediumSoftware EngineerTechnical ScreenSystem Design

Last updated

This problem tests your ability to design a multi-stage recommendation pipeline that balances low-latency retrieval with high-precision ranking at scale.

Scenario

Design a personalized content feed system (e.g., similar to Facebook News Feed or Instagram Explore) that recommends items to users in real-time. The system must serve 100 million daily active users with low latency and handle a massive volume of content updates.

Requirements

Functional Requirements

  • Provide a personalized list of content items for a user.
  • Support real-time updates: new posts should be eligible for recommendation shortly after being published.
  • Provide an API endpoint get_feed(user_id, count).

Non-Functional Requirements

  • Latency: The system should return recommendations in under 200ms.
  • Scale: Support high read and write throughput.
  • Freshness: Minimize the time-to-feed for newly created content.

Evaluation Criteria

  • Candidate Generation: How do you narrow down millions of items to a few hundred candidates?
  • Ranking: How do you score these candidates for a specific user?
  • System Architecture: How do the components (Retrieval, Ranking, Feature Store) interact?
  • Feedback Loops: How do you incorporate user interaction data (clicks, likes, skips) to improve the model?

Constraints

  • Assume you have an existing ML model service that can score a candidate list, but it is too slow to run on the entire corpus.
  • The system must handle cold starts for both new users and new content.

Hints

Hint 1

Consider a funnel approach: use fast, lightweight retrieval methods to filter the corpus, followed by a heavier, more accurate machine learning model to rank the results.

Hint 2

Think about how to balance 'exploitation' (showing content similar to what the user likes) with 'exploration' (showing new, potentially interesting content to avoid filter bubbles).

Hint 3

Identify where caching can be applied, noting the trade-offs between serving pre-computed feed results and calculating them on the fly.

Follow-up questions

  • How would you handle a 'flash crowd' event where a specific piece of content suddenly goes viral and requires massive scaling?
  • How would you measure the performance of your system using offline metrics (like AUC or NDCG) vs. online metrics (like CTR or dwell time)?
  • How does your design change if you need to enforce strict diversity requirements, such as ensuring content from different categories is displayed?