At KubeCon in Salt Lake City the CNCF announced yet another certification, the Certified Kubernetes Network Engineer (CKNE). Of course it is “for the AI era”, because what isn’t these days. Still, if you ask me, this is one of the more useful additions to the ever growing zoo of Kubernetes certificates (if you collect enough of them you become a “Kubestronaut”, I’m not making this up).
Like the CKA it is a hands-on exam where you have to fix things in a live environment instead of ticking multiple choice boxes. The topics are what you’d expect: CNI, services and DNS, traffic management, network policies and observability. Details are on the exam page.
Why do I think it’s a good idea? Because networking is where most Kubernetes clusters get hairy. Deploying a pod is easy. Finding out why that pod can’t talk to another pod on a different node, or why a service resolves in one namespace but not in the other, is a different story. Usually it ends with somebody staring at nftables rules generated by kube-proxy that no human being was ever supposed to read.
It’s all software defined now
What a lot of developers don’t realize is that a Kubernetes cluster is a software defined network. The pod network is typically an overlay (VXLAN or Geneve) on top of the physical network, routing decisions are made by a CNI plugin, policies are enforced in the kernel, more and more via eBPF (Cilium and friends) instead of iptables. The control plane that decides where packets go is completely separated from the hardware that actually moves them. That was the big promise of SDN back in the OpenFlow days, and while OpenFlow itself never really took over the world, the idea did. Today it just comes as YAML.
The upside is that you can build, change and tear down networks with the same tooling you use for everything else. The downside is that there are now several layers of abstraction between your packet and the cable, and when something breaks you need to understand all of them. Knowing your way around kubectl is not enough; you need to know what a veth pair is, how ARP works and why MTU suddenly matters when you wrap packets in other packets.
Data center networking strikes back
For years networking was the boring part of the data center. Leaf-spine fabric, BGP, done, let the cloud provider handle the rest. With AI workloads this changed quite dramatically. Training a large model means thousands of GPUs exchanging huge amounts of data all the time and the slowest link determines how fast the whole job goes. Suddenly people care about things like RDMA, RoCE, lossless Ethernet, congestion control and whether to use InfiniBand or not. The network is no longer a commodity, it is a significant part of what makes a GPU cluster fast (or expensive and slow).
Nvidia saw this coming early. In 2019 they announced the acquisition of Mellanox, back then best known for InfiniBand and high-speed network cards. So the company that sells you the GPUs now also sells you the NICs, the switches and the cables to connect them. One part of that portfolio is the BlueField DPU (Data Processing Unit): a network card with its own Arm cores that takes over networking, security and storage work from the host CPU. Things like virtual switching, encryption or firewalling run on the card itself, so the SDN data plane I was talking about above literally moves into a separate piece of hardware. That’s great for performance and for isolation between tenants, but it also adds yet another layer you need to understand when you’re debugging why your packets are going nowhere.
That’s also where the Kubernetes world and the classic network engineering world collide. The platform engineers know Kubernetes but think a switch is something you flip, the network engineers know their fabric inside out but get nervous when you mention a sidecar. Somebody has to bridge that gap, and a certification that forces people to actually understand the networking stack of Kubernetes is a step in the right direction.
Whether you need a certificate to prove that is another question. But having a curriculum that tells you what you should know is valuable in itself, even if you never take the exam. I personally learned most of my networking by breaking my home network repeatedly, which is cheaper but not quite as good for the CV.
