Adapt to Unified and Pluggable Cluster Management Platform at LinkedIn
summary
LinkedIn moves RAIN (their cluster resource manager) toward a Kubernetes-based unified platform, running across tens of thousands of hosts per cluster and multi‑cloud including Azure. They validate Kubernetes with a 160‑node experiment, integrate JupyterHub, TF Operator and TF Serving, and implement Kerberos delegation tokens for HDFS access. Real fixes include fixing an automation that unset ipv4_forwarding and upgrading to Kubernetes 1.14 to resolve init‑container env injection enabling secure, GPU‑backed ML training and serving.
AI-written summary · about
chapters (5)
- 0:00Start
- 0:42Agenda
- 3:49Typical Manual Deployment Workflow
- 9:53Jupiter Hub on Kubernetes
- 22:10Serving Serving Models Directly from Our Kubernetes Cluster
related talks
29mAdapt to Unified and Pluggable Cluster Management Platform at LinkedIn
40mFrom Metal To Apps: LinkedIn’s Kubernetes-based Compute Platform
18mKeynote: Adapting Kubernetes for Machine Learning Workflows
42mThe Challenges Managing a Kubernetes-Based Machine Learning Infrastructure
20mHow LinkedIn Performs Maintenances at Scale
20mReliable Migration Of 4+ Million CPU Cores of Compute to Kubernetes
33mKubernetizing Big Data and ML Workloads at Uber
53mImplementing Kubernetes on with confidence