Pinterest·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say

Jun 2026

Summary

Pinterest ML system design round, focused entirely on building a recommender system end to end. The depth they wanted on infrastructure surprised me, it wasn't just 'here's a two-tower model' and done.

Questions Asked (1)

Design a recommender system from scratch. Walk through where you'd store training data, features, and embeddings, how you'd handle offline versus online feature stores, and how you'd serve both the retrieval and ranking models at scale including model registry, low-latency inference, and traffic experimentation.

System DesignTechnical Trade-offsA/B Testing & Experimentation

Author's notes

This is a lot to cover in one question and I didn't pace myself well.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the scope and requirements, then structure your answer around the end-to-end ML lifecycle: data storage, feature engineering, model training, deployment, and experimentation. Focus on the two-stage retrieval and ranking architecture, explaining how you would handle offline vs online features, low-latency serving, and A/B testing at Pinterest scale.

Pro tip: Emphasize the importance of feature consistency between training and serving, and discuss how you would monitor and mitigate training-serving skew. Also, highlight the need for a robust experimentation platform to iterate quickly and measure business impact.

1. Clarify Requirements and Scope

Ask questions to understand the scale, latency requirements, and business goals. Confirm the two-stage retrieval and ranking approach and the need for real-time features.

2. Data Storage and Feature Management

Describe where training data, features, and embeddings are stored (e.g., data lake, feature store, vector database). Explain offline vs online feature stores and how to ensure consistency.

3. Model Training and Registry

Outline the training pipeline, including how models are versioned and stored in a model registry. Discuss how to manage model artifacts and metadata for reproducibility.

4. Serving Architecture for Retrieval and Ranking

Explain how to serve retrieval (e.g., ANN search) and ranking models with low latency. Cover model deployment, scaling, and caching strategies.

5. Experimentation and Monitoring

Describe how to run A/B tests, manage traffic, and monitor model performance and system health. Include feedback loops for continuous improvement.

Key Points to Mention

Two-stage architecture: retrieval (candidate generation) and ranking (personalization)
Feature store for offline/online consistency and low-latency feature serving
Vector database for embedding storage and approximate nearest neighbor search
Model registry for versioning, metadata, and deployment
Low-latency inference: model optimization, caching, and horizontal scaling
A/B testing framework for experimentation and guardrail metrics

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.