Building a Nodeless Kubernetes Platform
summary
Google builds a “nodeless” mode called GKE Autopilot that keeps real node objects visible but removes the VM API, uses cluster autoscaler and node auto-provisioner to create node pools from pod specs, and enforces limits with mutating and validating admission webhooks. It supports pod sizes from 0.25 to 28 cores, charges by pod request, allows certain daemonsets via allow‑listing, and reports a node‑creation latency of about 60–80 seconds; the platform is operational and production-ready.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:40Agenda
- 1:13Audience Poll
- 3:04Position Paper
- 4:00Kubernetes
- 5:32Operational Models
- 8:20Visibility
- 10:36Implementation
- 13:59Additional Components
- 15:58enforced policies
- 18:07why do we need it
- 20:47the result
- 22:01the benefits
- 27:07infrastructure features
- 29:26bursting
- 31:05downsides
- 32:44summary
- 33:40QA
related talks
35mKubernetes the easy way
20mKubernetes and GKE for Beginners
40mKubernetes - Easy Peasy
46mBuilt for billions, tailored for games; harnessing Google for serving players
16mRunning Services at Google Scale
20mKnative: A Kubernetes Framework to manage Serverless Workloads
31mKubernetes and the dynamic world in the cloud
07mKeynote: High Performance Computing on Google Kubernetes Engine