Katkat
Back

Design a Short-Video Recommendation System

MetaMachine Learning Engineer · Technical Screen
MediumMachine Learning EngineerTechnical ScreenML System Design

Last updated

This question tests your ability to design a multi-stage recommendation pipeline for a high-traffic short-video platform, focusing on candidate retrieval, ranking models, and system scalability.

Problem Statement

You are tasked with designing the backend recommendation engine for a new short-video platform similar to Instagram Reels or TikTok. The system needs to serve personalized content to millions of daily active users with sub-100ms latency.

Functional Requirements

  • Provide a personalized feed of videos to users.
  • Support real-time feedback loops (likes, watch time, skips).
  • Handle both "hot" viral content and niche interest-based discovery.

Non-Functional Requirements

  • Extremely low latency (e.g., <100ms for retrieval and ranking).
  • High throughput: millions of requests per second.
  • Scalability: System must handle both new creators (cold start) and viral spikes.

Design Deliverables

  1. High-level architecture: Describe the stages of the recommendation funnel (Candidate Generation, Ranking, Re-ranking).
  2. Model Selection: What models would you use for each stage and why?
  3. Data Strategy: How do you handle user/video features and real-time streaming data?
  4. Trade-offs: Address the exploration-exploitation dilemma and latency versus accuracy.
  5. Evaluation: How do you measure the success of the system both offline and online?

Hints

Hint 1

Consider a funnel approach: start with a large candidate pool and apply increasingly sophisticated models to smaller subsets.

Hint 2

Think about how to balance 'exploit' (what the user likes) with 'explore' (what the user might discover) to prevent filter bubbles.

Hint 3

Address the cold-start problem for new videos: how can you gather signals without relying solely on historical engagement?

Follow-up questions

  • How would you modify your ranking architecture to incorporate cross-modal signals, such as audio content or video frames?
  • If a specific creator suddenly goes viral, how does your system ensure the database and cache layers remain performant?
  • What specific metrics would you monitor to detect if the recommendation quality is degrading over time due to model drift?