Powering Spotify's Audio Personalization Platform
summary
Claims concrete scale and working components: personalization for 400 million monthly users across 183 markets over an 85+ million catalog, ~3 million predictions per second and hundreds of thousands of pipeline runs yearly. Infrastructure names and roles appear: Hendrix platform for ML productivity, Salem for model deployment and serving, jukebox for feature storage and reuse, and kubernetes‑based Q flow plus TFX for training pipelines. Examples include continuous training tradeoffs and measurable loop speedups, not just vague promises.
AI-written summary · about
chapters (8)
- 0:00Personalization and ML at Spotify
- 25:38Hendricks ML platform mission
- 26:44Model deployment with Salem
- 28:21Feature engineering with Jukebox
- 29:36Model training with Kubeflow
- 30:39Observability and metadata
- 32:07Impact of infrastructure
- 34:05Summary and Q&A
related talks
29mHow Spotify Does ML At Scale
30mTeam Teardown: Driving Platform Adoption and Success
27mPlay
27mEmpowering Traceable & Auditable ML in Production
42mThe Challenges Managing a Kubernetes-Based Machine Learning Infrastructure
30mSpotify Harnesses Ray for Next-Gen AI Infrastructure
38mBuilding and Managing a Centralized Kubeflow Platform at Spotify
32mScaling Kubeflow for Multi-tenancy at Spotify