⚡ Lightning Talk: Data Science Environments in Seconds: Scaling Jupyter Notebooks in Kubernetes
Lightningat AppDeveloperCon NA 2024 · on Tuesday, 2024-11-12· at 12:25· Salt Palace | Level 1 | 151 G·
summary
Pre-warmed Jupyter kernels run in Kubernetes so kernels go from minutes to seconds, claiming a 96% latency reduction by keeping standby kernels ready to attach to Jupyter servers. Kernel specs are ranked by recent launch frequency (past ~3 days) and eligible namespaces are auto-enabled. Pools use small initial resources with dynamic Spark executor scaling and backend cycles that scale up/down; pool size stays under pool <10% of namespace quota and tear down long-idle kernels.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:04Jupyter environments context
- 3:12Kernel launch delays
- 3:55Pre-warmed kernel pools
- 5:21Demo: pre-warmed kernels
- 6:10Performance and resources
- 8:42Scaling policies
- 10:14Future improvements
- 11:40Conclusion
related talks
39mHow We Hosted 200+ Data Scientists in a K8S Cluster
34mMoving notebooks into the cloud: challenges and lessons learned
44mEmpowering Data Science with Kubernetes | Samuele Chiocca and Andrea Rota
24mJupyter Notebooks for Ops
49mExperimenting with Large Data Volumes Using Notebooks
32mSpin up Jupyter notebooks at scale and enhance your productivity
15mKeynote: Machine Learning on Kubernetes Made Easy With Kubeflow
28mDevOps for Data Science With Kubernetes