How should I approach writing a simple CNI plugin for learning?

Start with the basics before attempting anything like Cilium. On a single node, make the plugin assign a pod IP and configure the necessary routes. Then add a small IPAM implementation. Once that works, expand to multiple nodes using static routes between nodes. After you understand that model, explore eBPF, starting with traffic-control hooks.

The CNI interface itself is fairly small, so a barebones implementation is a realistic learning project. Begin by reading the CNI specification and implementing the lifecycle operations for adding and removing a network. A simple plugin can create a veth pair and configure the pod side, while a separate node agent handles routes or tunneling between hosts. Looking at a small VXLAN-based project such as Flannel can also show how the inter-node portion fits together. Once the fundamentals are clear, move on to eBPF and kernel-level packet processing.

It helps to separate the project into two networking problems. First is pod-to-host connectivity: the CNI executable runs when a pod sandbox is created or removed, and can create a veth pair, place one end in the pod network namespace, assign addresses, and configure routes. Second is inter-node connectivity: a process running on every node needs to connect pod networks across hosts, commonly with an overlay such as VXLAN. A small overlay-based implementation is much easier to understand than jumping straight into all of Cilium’s features.