Co-Evolution of Kubernetes and GCP Networking
summary
Kubernetes shifts from node NATs to giving pods real IPs by using alias IPs, removing host port collisions; GCP scales route-based prototype from routes 50→5,000 and then replaces it with Network Endpoint support so load balancers target pods directly via Network Endpoint Groups, fixing double‑hop and cookie‑affinity breakage. Services map VIP/nodePort semantics; GCLB integration uses iptables initially (fast) but external LB programming is slower causing rollouts to fail until readiness gates let load balancers signal readiness during updates.
AI-written summary · about
chapters (37)
- 0:00Intro
- 2:44Networking
- 3:26Kubernetes
- 4:28Scheduling Pass
- 5:30Docker
- 6:33Voluntary cooperative sharing
- 7:15Kubernetes model
- 8:31Direct flat network
- 8:41Prototyping
- 9:21IP Routing
- 11:36The Catch
- 13:04GK Cluster
- 13:18PodRangers
- 13:32Pod Ranges
- 15:02Services
- 15:56VIP Load Balancers
- 17:57Noda Port
- 18:29Difference between the two models
- 18:58Design of the communities service model
- 20:09L7 load balancers
- 20:56Ingress
- 22:00IP Tables
- 23:51Cookiebased affinity
- 24:16Packet flow
- 25:08Load balancer
- 25:26Network one groups
- 26:24TCP networking evolved
- 26:42Programming external load balancers
- 29:55Cloud providercentric Kubernetes
- 31:18Backend Conflict
- 31:56Recap
- 32:17Too Flexible
- 33:23Not Flexible Enough
- 33:49Service Abstraction
- 34:26Summary
- 35:31Intro and Deep Dive
- 36:00Questions
related talks
34mLife of a Packet [I]
32mCruise’s Self-Driving Networking Journey
20mKubernetes Networking 101
51mKubernetes SIG-Network: Intro and Deep-Dive
40mCloud, Kubernetes, and Service Networking - Taming the Turtles
09mLightning Talk: Hacking Kubernetes Networking to Create Telco Solutions
19mSIG-network: Updates and Future Directions
40mKubernetes Networking Panel: Collaborate on the Spec, Compete on Execution