Building Recommendation Systems for Global E-commerce Marketplaces
summary
Adevinta builds centralized recommender and profiling services using Vertex AI and GCP (BigQuery, Spark, Cloud Run, Cloud Functions), migrating from on‑prem to cloud and managing infra with Terraform and Atlantis. Systems serve about 7,000 requests per second with a P99 100ms API and reach 120 million monthly users; recommendations lift leads ~3–5%. Development uses Vertex AI notebooks and pipelines for experimentation, then Airflow in production; pipelines run daily, publish to BigQuery, and monitoring/alerts detect failures.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:50Adevinta and its marketplaces
- 3:38Personalization team overview
- 7:02Recommendations with complementary items
- 8:24ML lifecycle phases
- 10:34ML platform guiding principles
- 12:22Google Vertex AI adoption
- 14:49Terraform and Atlantis infrastructure
- 15:52Vertex AI notebooks
- 17:20BigQuery data pipelines
- 19:06Vertex AI Pipelines orchestration
- 21:53Recommendations via BigQuery and email
- 23:40Initial recommendation pipelines
- 24:50Vertex AI Pipelines challenges
- 27:18Cloud Composer (Airflow) switch
- 29:28Project guidelines and templates
- 30:30Key takeaways
related talks
32mResurrecting a Recommendations Platform
29mBuilding ML Platform Capabilities at Global Scale
33mPowering Real-time Recommendations with Machine Learning on GCP
28mThe Power of Personalized Customer Experiences Through AI Innovations
39mDeveloping Machine Learning for Impact
29mDeveloping Visual AI Solutions for Online Marketplace
29mThe Challenges of Building a Personalized Recommendation System in E-Commerce
24mAI-Powered Transformation in Fashion E-Commerce