/packet-ipv6-kubernetes

Primary LanguageNixApache License 2.0Apache-2.0

IPV6 cluster on Packet

Sets up IPV6-native cluster using packet.

This shows how a modern container fabric should look like. Without any NAT or overlay nonsense; but a pure IP/Layer-3 networking fabric.

Instead of using an overlay network, each pod is directly assigned an IPV6 address. Routes to pods are announced directly to the packet router BGP.

Furthermore ClusterIPs are also allocated from a public range. This means we can publically announce kubernetes services directly; without the need of dedicated cloud load-balancers! Because Packet's routers support ECMP; traffic will be load-balanced between nodes that announce the service automatically.

The IPV6-native world is a bit different than you're used to. Normally in k8s we kind of rely on NAT and overlaynetworks for a false sense of security, and hence most people don't even set up NetworkPolicy's for their services and pods properly. However, with IPV6 there are no overlays and NAT; so we are forced to take security seriously and explicitly define NetworkPolicies to make sure only the things that should talk to eachother can talk to eachother.

Usage

Create an .envrc.local with your secrets:

export TF_VAR_kubeadm_certificate_key=
export TF_VAR_kubeadm_token=
export PACKET_AUTH_TOKEN=

$ direnv allow
$ apply

apply might fail to get the kubeconfig on first try; as the cluster might still be bootstrapping. Type apply again to get it. The script is idempotent. It is important to make run it until it succeeds as this also configures the BGPPeers with the packet routers.

Stream notes

https://www.youtube.com/watch?v=ohwHzr4O6Fo

Gotchas

  • Enable ipv6 forwarding
  • Set up ipv6 dns servers on host as coredns will use the host's dns servers
  • Docuemtnation for both kubeadm and calico incomplete or outdated:
    • need to set kubelet node-ip
    • need to disable ipv4 autodetection on calico

Packet, pods, ipv6

packet

  • Pod network is flat. all pods should reach eachother without NAT
  • pod cidr != node cidr != service cidr
  • Overlay networks (usually; but not neccessarily; E.g. GCP sets up routes directly for pods)
  • Pod network usually a private range. vxlan
  • What if each pod just got a public ip address? => No need for overlays
  • Very possible with IPv6 (Packet gives you a /56)
  • Packet allows us to announce routes using BGP.
  • Calico can announce pod ip addresses to BGP automatically
  • https://docs.projectcalico.org/reference/architecture/design/l3-interconnect-fabric
  • downward default network downward

Advertising ClusterIPs

  • We can directly advertise kubernetes services using BGP too!
  • Will use ECMP routing to loadbalance between nodes. So highly available LB
  • Just make sure that the ip range is also a public ipv6 range
  • No need for external load-balancer; or metallb, or whatever! All services are reachable through public IP Address

externalTrafficPolicy: Local vs Cluster

  • Show what it means for routes

bbbbut What about security?! (Out of scope probably due to time constraints!)

Whats next - Security