Scaling Kubeflow for Multi-tenancy at Spotify
Talkat KubeCon + CloudNativeCon North America 2021 · on Friday, 2021-10-15· at 17:25· 411 Theater + Online·
summary
Spotify runs Kubeflow Pipelines on GKE with a Python SDK and TFX, serving 60+ teams and ~600 users, logging 30,000 models and 100,000 pipeline hours (~300 runs/day). Platform uses Terraform + kustomize for cluster and namespace provisioning, a custom controller that maps team profiles via Google groups to Kubeflow namespaces, and a multi-cluster versioning strategy (per-SDK major-version clusters) to avoid breaking changes. Deployments use Argo CD and CI, observability adds SLOs in GCP and a custom kubeflow-state-metrics exporter for product-specific Prometheus metrics.
AI-written summary · about
chapters (28)
- 0:00Intro
- 1:48Kubeflow Platform for ML
- 2:28ML Workflow @ Spotify
- 4:48Kubeflow clusters
- 5:30Kubeflow platform stats
- 6:15Our solutions
- 6:58Team based multi-tenancy
- 9:03Team profile management
- 10:37Team profile example
- 12:49Existing setup & problems
- 14:12Multi-cluster strategy illustration
- 15:09Benefits of multiple clusters
- 15:57Multi-cluster implementation
- 16:53Multi-cluster based Kubeflow Platforr
- 17:48New challenges
- 18:53Solution: ArgoCD
- 19:36Multi-cluster CD
- 20:39Multi-cluster reliability
- 21:34"Infrastructure"-as-code
- 23:36SLO Tracking
- 24:45Telemetry and Metrics
- 25:34kubeflow-state-metrics
- 26:46Infrastructure and Infrastructui
- 28:18Product Identity
- 29:22Expanded Observability
- 29:56Expanded "On-Cluster" Compute
- 30:40Self-Service
- 31:14Open Source!
related talks
38mBuilding and Managing a Centralized Kubeflow Platform at Spotify
34mKubeflow: Multi-Tenant, Self-Serve, Accelerated Platform for Practitioners
30mTeam Teardown: Driving Platform Adoption and Success
54mScaling Machine Learning Pipelines in Cloud
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
51mMachine Learning using Kubeflow and Kubernetes
31mKubeflow - Data Science on Steroids
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
