Machine Learning on Kubernetes at Shell: A Kubeflow Journey
Talkat KubeCon + CloudNativeCon North America 2020 Virtual · on Wednesday, 2020-11-18· at 15:50· Intrado Virtual Event Platform,·
summary
Shell runs Kubeflow with Arrikto Rok on EKS using GitOps, Kale and KFP to turn notebooks into pipelines, spawn hundreds or thousands of hyperparameter runs, and keep ML metadata in MLMD. The stack snapshots notebook volumes about every 10 minutes for end-to-end reproducibility, archives to S3, and syncs across regions; claims 10,000 models in 2 hours, with local NVMe giving 16× read IOPS 400× write IOPS and 18–21× bandwidth versus EFS, and uses OIDC SSO + Istio/Envoy for enterprise security.
AI-written summary · about
chapters (17)
- 0:00Start
- 0:30WARNING: uncertainties ahead
- 1:14Shell Global
- 4:57Challenges, Solutions & Lessons Learned (1/2) Arrikto
- 7:37Challenges, Solutions & Lessons Learned (2/2) Arrito
- 9:50GitOps - Infrastructure as Code
- 11:19Gitops - Upgrade (2/2)
- 11:47Storage
- 13:23Rok - Run At The Speed of Flash
- 13:58Rok Cluster
- 14:08Rok Registry
- 14:49Rok - Across locations
- 15:26Security & Isolation (1/2)
- 17:05Data Science Workflows
- 18:44Why choose Kubeflow
- 20:54Demo
- 27:45Conclusion
related talks
51mMachine Learning using Kubeflow and Kubernetes
15mKeynote: Machine Learning on Kubernetes Made Easy With Kubeflow
46mDeploying scalable Machine Learning models with Kubernetes
54mScaling Machine Learning Pipelines in Cloud
14mKubeflow: Where ML Meets Kubernetes
54mEconomics and Best Practices of Running AI/ML Workloads on Kubernetes
39mDeploy your ML Workloads on Kubernetes with Kubeflow
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow