Keynote: Machine Learning on Kubernetes Made Easy With Kubeflow
summary
Kubeflow powers a self‑service ML platform that lets data teams launch notebook in minutes using preconfigured images (TensorFlow, PyTorch, Spark, GPU variants), ties notebooks into reproducible pipelines via Kale converts notebooks to pipelines, and reclaims idle capacity with a snapshot after 24 hours policy. Shell reports bringing a proof‑of‑concept from local prototype (1–2 months) to production in 35 days and subsequent projects down to a couple of days, showing the setup actually cuts turnaround time and cloud costs.
AI-written summary · about
chapters (11)
- 0:00Kubeflow and ML on Kubernetes
- 1:52Why Kubernetes and Machine Learning Fit Together
- 3:36ML model deployment challenges
- 4:19Kubeflow in ML workflows
- 5:02Shell’s Kubeflow journey
- 6:26Shell’s ML challenges
- 8:36Kubeflow solved infrastructure issues
- 10:23Live demo: notebook server
- 12:06Kubeflow boosts ML productivity
- 13:29Automating ML pipelines
- 14:13Conclusion & Q&A
related talks
10mKeynote: Machine Learning using Kubeflow and Kubernetes
22mKeynote: Kubeflow ML on Kubernetes
14mKubeflow: Where ML Meets Kubernetes
51mMachine Learning using Kubeflow and Kubernetes
30mKubeflow, Your ML Tools on Kubernetes
54mEconomics and Best Practices of Running AI/ML Workloads on Kubernetes
28mMachine Learning on Kubernetes at Shell: A Kubeflow Journey
39mDeploy your ML Workloads on Kubernetes with Kubeflow