How Kubernetes Networking Really Works: A Packet’s Journey Across Pods and Nodes
summary
Live demos build a kernel‑level mental model: a hand‑rolled tun + user‑space encapsulator works but loses ~25% of packets and adds 43 bytes per packet plus CPU and copy overhead, while replacing it with a kernel VXLAN device removes the user‑space bottleneck and gives 100% delivery using kernel VXLAN. ARP broadcasts and spoofing motivate moving from L2 bridges to host‑gateway /32 routing and finally BGP route advertisement (bird) to make pods natively routable.
AI-written summary · about
chapters (18)
- 0:00Kubernetes networking fundamentals
- 1:03Why we built this talk and CNI basics
- 3:51Linux primitives for networking
- 5:56Setting up pod networking
- 9:27Pod-to-pod communication demo
- 12:21Pod-to-pod across nodes challenges
- 16:43Encapsulation with tun device demo
- 20:59Asymmetric network topologies problems
- 23:03Eliminating L2 switching with L3 routing
- 26:13L3 routing without bridges demo
- 29:08User-space encapsulation limits
- 31:11VXLAN kernel encapsulation
- 35:56VXLAN demo
- 40:09Teaching the network about pods with BGP routing
- 41:32BGP demo
- 47:01Networking strategies trade-offs
- 48:25Kubernetes networking principles
- 50:11Summary and Q&A
related talks
20mKubernetes Networking 101
29mInside Kubernetes Networking
34mLife of a Packet [I]
09mLightning Talk: Hacking Kubernetes Networking to Create Telco Solutions
27mDeconstructing K8s Networking for Knowledge and Profit
40mAttacking Kubernetes Clusters Through Your Network Plumbing
32mThe Easy--Don't Drive Yourself Crazy--Way to Kubernetes Networking
17mPod-to-Pod Ping: The Hidden Wiring of Kubernetes