Design a YouTube-style video recommendation system
Company: Yahoo
Role: Machine Learning Engineer
Category: ML System Design
Difficulty: easy
Interview Round: Onsite
Design a recommendation system for a YouTube-like video platform. When a user opens the home page, or is watching a video, the system returns a ranked list of videos to show. The report gives only the name of the problem, so the surfaces, scale, objective and latency are open: settle them with the interviewer before designing.
```hint Count what you can score
Ask how many videos a heavy model could score for one request within the latency budget, compared with the size of the catalog.
```
```hint What a click misses
Think about what a model trained only to predict clicks would learn to recommend, and which other signals show that a user valued a video.
```
### Clarifying Questions
- Which surface: the home feed, the list of videos to watch next beside the player, or both?
- How many daily active users, how many videos in the catalog, and how many new uploads per day?
- What is the business objective: watch time, user satisfaction, long-term retention, or a mix? Are there goals on the creator side?
- What latency budget does a recommendation request have?
- Which signals exist: watch history, searches, subscriptions, likes, dismissals, surveys?
- Which policy constraints apply: content safety, freshness, diversity, or limits on videos from the same creator?
### What a Strong Answer Covers
- Requirements, scale estimates, and an objective that rewards real satisfaction rather than clicks alone
- A multi-stage design: candidate retrieval from several sources, ranking, and re-ranking with business rules
- Labels, features and training data, including negatives, position bias and delayed feedback
- Offline and online evaluation with guardrail metrics
- Serving within the latency budget, freshness of models and embeddings, and cold start for new users and new videos
- Monitoring, feedback loops, diversity and exploration
### Follow-up Questions
- A video was uploaded ten minutes ago. How does it get its first impressions, and how quickly can the system learn whether it is good?
- How do you keep recommendations from narrowing into a filter bubble?
- Training data comes only from what the current system chose to show. How does that bias the next model, and how do you correct for it?
- How would you use what the user watched in the last few minutes of the current session?
Overview: Design a recommendation system for a YouTube-like video platform that returns ranked videos for the home feed or the watch-next list. Tests choosing an objective beyond clicks, multi-stage retrieval and ranking, labels and bias, evaluation, low-latency serving, cold start and monitoring.
Read the full Yahoo Machine Learning Engineer interview experience this question came from