Katkat
Back

Design an E-Commerce Recommendation System from Browsing History

OracleSoftware Engineer · Onsite
MediumSoftware EngineerOnsiteML System Design

Last updated

This problem tests the ability to architect a scalable, low-latency recommendation engine that translates real-time user activity into personalized product suggestions.

Problem Statement

Design a recommendation system for an e-commerce platform that suggests relevant products to users based on their real-time browsing history. The system must handle high traffic volumes and update suggestions as a user navigates through the site.

Functional Requirements

  1. Capture and process user clickstream data in real-time.
  2. Provide personalized product rankings (top-K) for a specific user ID.
  3. Support "Related Items" suggestions based on the current product page view.

Non-Functional Requirements

  1. Latency: Recommendation requests must return in under 100ms.
  2. Scale: Support 10,000 requests per second (RPS).
  3. Availability: The system should remain functional even if the primary recommendation model is temporarily down.

Deliverables

Your response should cover the following:

  • Data Flow: How do you ingest clickstream data, and how do you store user profiles/history?
  • Model Architecture: Choose a candidate generation and ranking strategy (e.g., collaborative filtering vs. content-based vs. embeddings).
  • API Design: Define the necessary endpoints for the frontend to consume.
  • Storage: Decide on the database choice for user features and product metadata.
  • System Architecture: Propose a high-level design diagram, including caching strategies and potential trade-offs between accuracy and latency.

Hints

Hint 1

Consider a two-stage approach: candidate generation (retrieving a large set of potentially relevant items) followed by a fine-tuned ranking layer.

Hint 2

Think about using an asynchronous message queue (e.g., Kafka) to decouple real-time data ingestion from model inference.

Hint 3

Address the cold-start problem: how do you handle users or products with no historical data?

Follow-up questions

  • How would you handle a sudden surge in traffic during a seasonal sale (e.g., Black Friday)?
  • How do you evaluate the performance of your recommendation engine using offline metrics versus online A/B testing?
  • How would you ensure user data privacy while maintaining the personalization quality of the system?