How Spotify Does ML At Scale
summary
Spotify reports building an ML platform since 2017 that centralizes tooling—Jukebox feature store, Kubeflow Pipeline (skf/qflow), Salem model serving and ML Home—backed by BigQuery/GCS/Bigtable and TFRecord pipelines. Teams train ~30,000 models in 2020, serve ~300,000 predictions/sec, and see model-iteration velocity up 700%. The platform reduces time-to-production and tech debt, uses GPUs for audio and some training, ships a sing-along media feature, and still struggles with early prototyping and centralized governance.
AI-written summary · about
chapters (6)
- 0:00Introduction and welcome
- 2:12Why we built ML platform
- 8:13Platform goals and components
- 14:34Deep dive into Jukebox
- 18:23Key learnings and future plans
- 22:55Q&A session
related talks
40mPowering Spotify's Audio Personalization Platform
38mBuilding and Managing a Centralized Kubeflow Platform at Spotify
26mManaging ML Models @ Scale
27mEmpowering Traceable & Auditable ML in Production
13mRevolutionizing the Music Industry with Data Science
32mScaling Kubeflow for Multi-tenancy at Spotify
42mThe Challenges Managing a Kubernetes-Based Machine Learning Infrastructure
23mML Platform Challenges & Applications at Freshworks