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

Senior

Jul 2026

Summary

Pinterest ML engineer interview focused entirely on a large-scale URL health detection system. One long design question that touched basically every layer of the stack, from crawling strategy to downstream ad system integration. Dense but interesting.

Questions Asked (1)

Design an ML system to detect broken or low-quality landing pages (404s, soft-404s, parked domains, redirect loops) across a massive URL inventory. Walk through labeling, features, crawling strategy, model architecture, serving, and success metrics.

System DesignTechnical Trade-offsProduct Analytics & Metrics

Author's notes

This question is basically six questions stitched together and they want you to move through all of them without getting stuck.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the scale and constraints, then walk through the ML lifecycle: data labeling, feature engineering, crawling strategy, model selection, serving, and metrics. Emphasize trade-offs between precision and recall, and how to handle the massive scale with efficient sampling and distributed processing.

Pro tip: Propose a two-stage system: a fast rule-based filter to catch obvious cases (404s, redirect loops) and a lightweight ML model for ambiguous cases (soft-404s, parked domains). This reduces computational cost and improves precision.

1. Clarify Requirements and Scale

Ask about URL inventory size, update frequency, latency requirements, and acceptable false positive/negative rates. This informs the entire design.

2. Data Labeling and Feature Engineering

Define labeling criteria (e.g., manual review, heuristics) and design features from URL, content, and metadata (e.g., HTTP status, redirect chains, text quality, domain age).

3. Crawling and Data Collection Strategy

Design a prioritized crawling system that samples URLs efficiently, respects robots.txt, and handles scale with distributed workers and caching.

4. Model Architecture and Training

Choose a model (e.g., gradient boosted trees for tabular features, or a hybrid with text embeddings) and address class imbalance with techniques like focal loss or resampling.

5. Serving, Monitoring, and Success Metrics

Deploy as a batch or streaming service, monitor drift, and define metrics like precision@k, recall, and business impact (e.g., reduction in user reports).

Key Points to Mention

Handling class imbalance and defining a clear labeling strategy with active learning.
Feature engineering: HTTP status codes, redirect chains, content entropy, domain parking signals, and text quality metrics.
Crawling at scale: distributed crawling, politeness policies, and incremental updates.
Model choice: trade-offs between interpretable models (GBDT) and deep learning for text.
Serving architecture: batch vs. real-time, and integration with existing pipelines.
Success metrics: offline (AUC, precision-recall) and online (user engagement, reduction in broken links).

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