From Metal To Apps: LinkedIn’s Kubernetes-based Compute Platform
Talkat KubeCon + CloudNativeCon Europe 2025 · on Thursday, 2025-04-03· at 11:45· Level 0 | ICC Capital Hall | Room I·
summary
LinkedIn runs Kubernetes on bare metal across 500,000+ servers and 1.5M containers, using a gRPC compute broker and inventory manager to allocate heterogeneous pools; a maintenance orchestrator stripes data centers into 20 maintenance zones so upgrades affect ≤5% capacity and deliver kernel and kubelet updates with no downtime. Clusters scale toward 5K nodes, users deploy via LI custom resources (LI deployment/statefulset) backed by clone sets, per-cluster Argo CD (migrating to an internal GitOps engine), a generic stateful controller for shard managers, and CPU pinning and IPv6 pod networking for isolation and performance.
AI-written summary · about
chapters (19)
- 0:00LinkedIn’s compute platform intro
- 2:48Infrastructure as a Service layer
- 7:42Kubernetes cluster management
- 11:14Workload platform overview
- 13:01Avoiding Kubernetes built-ins
- 14:02Stateful systems on Kubernetes
- 15:47Stateless app deployment
- 18:52Argo CD challenges at scale
- 20:16Failure handling automation
- 21:42Debugging tools
- 23:27Multi-cluster federation plans
- 25:14Migration from legacy systems
- 26:36Q&A session part one
- 29:06Q&A session part two
- 31:59Q&A session part three
- 33:47Q&A session part four
- 35:10Q&A session part five
- 36:58Q&A session part six
- 38:00Q&A session part seven
related talks
29mAdapt to Unified and Pluggable Cluster Management Platform at LinkedIn
29mAdapt to Unified and Pluggable Cluster Management Platform at LinkedIn
40mOperating a Self-Healing Bare-Metal Kubernetes Platform at Global Scale
14mKeynote: Kubernetes as a Foundational Layer of Infrastructure
23mMore Power, Less Pain: Building an Internal Platform with CNCF Tools
20mKeynote: Kubernetes at GitHub
36mBuilding a CI/CD driven infrastructure for managing Kubernetes clusters
22mKeynote: Cloud Native at Alibaba Scale