How Pinterest Rebuilt Its Recommendation Engine for 500 Million Users

Episode notes

In this episode of The CTO Podcast, Lucas and Luna dive into how Pinterest's engineering team rebuilt its core recommendation engine from a batch-processing pipeline to a real-time, graph-based system serving over 500 million monthly active users. They explore the specific architectural decisions Pinterest made: moving from collaborative filtering to a heterogeneous graph neural network called PinSage, deploying it on TensorFlow Serving with Kubernetes for low-latency inference, and handling the cold-start problem for new pins and users. The discussion covers the trade-offs between offline batch precomputation and online inference, how Pinterest reduced recommendation latency from hours to milliseconds, and the infrastructure costs involved. Lucas explains the graph-based approach that captures user intent through 'pins' and 'boards,' while Luna questions how this impacts content discovery and engagement. Tune in for a detailed look at a real-world scale challenge in modern machine learning systems.

• How Airbnb Rebuilt Search for 8 Million Listings62 / 100
• How GitLab Built a Single Codebase for One Million CI Pipelines65 / 100
• How Slack Rebuilt Its Search Index for 10 Million Daily Queries57 / 100
• How Notion Rebuilt Its Sync Engine for Offline-First
• How Notion Rebuilt Its Block Engine for Hybrid Local-Sync