Keynote: Building Enterprise AI/ML Platform on Kubernetes
summary
Organizations build ML platforms on Kubernetes for four reasons: portability, customizability, performance, consistency, claiming Kubernetes supports autoscaling from single nodes to thousands for distributed training and hardware accelerators like GPUs/TPUs. Recommended components include a feature store (Feast), Jupyter notebooks, TensorBoard, Kubeflow Pipelines, training operators and KServe for multi-model serving; use spot VMs with checkpointing or elastic training operators, Evidently for model drift, and MLflow for model registry. Kubeflow donated to CNCF.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:22Why do organizations build their platform on Kubernetes
- 2:13Building endtoend ML applications
- 3:14Machine learning is an experimental discipline
- 4:02Feature store Notebooks and Pipelines
- 5:28Constant Retraining
- 5:56Key Requirements
- 7:40Logging Monitoring
- 9:44Model Management
- 10:57Crewflow
- 11:20Wrap up
related talks
44mEngaging the Kubeflow Community: Building an Enterprise-Ready AI/ML Platform
22mKeynote: Kubeflow ML on Kubernetes
10mKeynote: Machine Learning using Kubeflow and Kubernetes
15mKeynote: Machine Learning on Kubernetes Made Easy With Kubeflow
08mWelcome + Opening Remarks
14mKubeflow: Where ML Meets Kubernetes
46mDeploying scalable Machine Learning models with Kubernetes
31mKubernetes for Machine Learning: Managing and Scaling AI Workloads