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

SeniorPrefer not to say

May 2026

Summary

ML system design round at Meta for an MLE role. The prompt was designing the recommendation page for Facebook Marketplace, which sounds scoped but opens up into a pretty deep rabbit hole once you start thinking about candidate retrieval, ranking, and cold start.

Questions Asked (1)

Design a machine learning recommendation system for the Facebook Marketplace feed.

System DesignTechnical Trade-offsProduct Sense & Ideation

Author's notes

Spent probably too long on the problem framing before getting into the actual ML components.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product goals and constraints, then outline a high-level system architecture covering data, features, models, and serving. Dive into key design choices like candidate generation, ranking objectives, and evaluation metrics, while discussing trade-offs and scalability.

Pro tip: Emphasize the two-stage ranking (candidate generation + ranking) and how you'd handle cold-start and real-time signals, as these are critical for Marketplace's dynamic inventory. Also, tie your design to Meta's business metrics like engagement and transactions.

1. Clarify Requirements and Goals

Ask questions to understand the product: What are the key objectives (e.g., increase transactions, engagement)? What are the constraints (latency, scale)? Who are the users (buyers, sellers)?

2. Outline High-Level Architecture

Describe the end-to-end system: data collection (user interactions, item features), feature engineering, model training, and serving. Mention offline vs. online components.

3. Design Candidate Generation and Ranking

Explain the two-stage approach: first, generate a set of candidate items (e.g., via collaborative filtering, content-based, or graph-based methods); then rank them with a more complex model (e.g., deep learning) predicting CTR or conversion.

4. Address Key Challenges and Trade-offs

Discuss cold-start for new users/items, real-time personalization, scalability, and how to balance exploration vs. exploitation. Mention trade-offs between model complexity and latency.

5. Define Evaluation and Metrics

Propose offline metrics (AUC, NDCG) and online metrics (CTR, conversion rate, engagement). Explain how to set up A/B tests and monitor for bias or feedback loops.

Key Points to Mention

Two-stage ranking: candidate generation (e.g., using embeddings, ANN) followed by ranking (e.g., DNN with features like user-item interaction, item popularity, recency).
Feature engineering: user features (demographics, past behavior), item features (price, category, images), and context features (time, location).
Cold-start solutions: use content-based features for new items, and for new users, leverage demographic or session-based signals.
Real-time personalization: incorporate streaming features (e.g., recent clicks) and update models frequently (e.g., online learning).
Evaluation: offline metrics like precision@k, NDCG; online A/B testing with business metrics (transactions, revenue).
Scalability: distributed training, model serving with low latency, and efficient candidate retrieval (e.g., FAISS for ANN).

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