This module is responsible for providing a network between multiple nodes in a cluster using the cilium module.

Limitations

  1. This module currently supports only direct-routing mode.
  2. Service types NodePort and LoadBalancer do not work with hostNetwork endpoints in the DSR LB mode. Switch to SNAT if it is required.
  3. HostPort Pods will bind only to one interface IP. If there are multiple interfaces/IPs present, Cilium will select only one of them, preferring private IP space.
  4. Kernel requirements.
    • The cni-cilium module requires a Linux kernel version >= 4.9.17.
    • For the cni-cilium module to work together with the istio, openvpn or node-local-dns module, a Linux kernel version >= 5.7 is required.
  5. OS versions support.
    • Ubuntu
      • 18.04
      • 20.04
      • 22.04
    • Debian
      • 11
    • CentOS
      • 7 (requires kernel from external repo)
      • 8 (requires kernel from external repo)

A note about CiliumClusterwideNetworkPolicies

  1. Make sure that you deploy initial set of CiliumClusterwideNetworkPolicies with policyAuditMode configuration options set to true. Otherwise you are degrading cluster operation or even completely losing SSH connectivity to all Kubernetes Nodes. You can remove the option once all CiliumClusterwideNetworkPolicy objects are applied and you’ve verified their effect in the Hubble UI.

  2. Make sure to deploy the following rule, otherwise control-plane will fail for up to 1 minute on cilium-agent restart. This happens due to conntrack table reset. Referencing kube-apiserver entity helps us to “circumvent” the bug.

    apiVersion: "cilium.io/v2"
kind: CiliumClusterwideNetworkPolicy
metadata:
  name: "allow-control-plane-connectivity"
spec:
  ingress:
  - fromEntities:
    - kube-apiserver
  nodeSelector:
    matchLabels:
      node-role.kubernetes.io/control-plane: ""

A note about Cilium work mode change

If you change the Cilium operating mode (the tunnelMode parameter) from Disabled to VXLAN or vice versa, you must restart all nodes, otherwise there may be problems with the availability of Pods.