infrastructure-documentation/k8s.md

#### Imaging

Using installimage from the rescue instance.

- `wipefs -fa /dev/nvme*n1`
- `installimage -r no -n hetzner0?`
- Debian bookworm
- `PART / ext4 100G`
- `PART /srv ext4 all`
- ESC 0 + yes
- reboot

Partitioning.

- First disk
  - OS
  - non precious data such as the LXC containers with runners.
- Second disk
  - a partition configured with DRBD

Debian user.

- `ssh root@hetzner0?.forgejo.org`
- `useradd --shell /bin/bash --create-home --groups sudo debian`
- `mkdir -p /home/debian/.ssh ; cp -a .ssh/authorized_keys /home/debian/.ssh ; chown -R debian /home/debian/.ssh`
- in `/etc/sudoers` edit `%sudo   ALL=(ALL:ALL) NOPASSWD:ALL`

#### Install helpers

Each node is identifed by the last digit of the hostname.

```sh
sudo apt-get install git etckeeper
git clone https://code.forgejo.org/infrastructure/documentation
cd documentation/k3s-host
cp variables.sh.example variables.sh
cp secrets.sh.example secrets.sh
```

Variables that must be set depending on the role of the node.

- first server node
  - secrets.sh: node_drbd_shared_secret
- other server node
  - secrets.sh: node_drbd_shared_secret
  - secrets.sh: node_k8s_token: content of /var/lib/rancher/k3s/server/token on the first node
  - variables.sh: node_k8s_existing: identifier of the first node (e.g. 5)
- etcd node
  - secrets.sh: node_k8s_token: content of /var/lib/rancher/k3s/server/token on the first node
  - variables.sh: node_k8s_existing: identifier of the first node (e.g. 5)
  - variables.sh: node_k8s_etcd: identifier of the node whose role is just etcd (e.g. 3)

The other variables depend on the setup.

#### Firewall

`./setup.sh setup_ufw`

#### DRBD 

DRBD is [configured](https://linbit.com/drbd-user-guide/drbd-guide-9_0-en/#p-work) with:

`./setup.sh setup_drbd`

Once two nodes have DRBD setup for the first time, it can be initialized by [pretending all is in sync](https://linbit.com/drbd-user-guide/drbd-guide-9_0-en/#s-skip-initial-resync) to save the initial bitmap sync since there is actually no data at all.


```sh
sudo drbdadm primary r1
sudo drbdadm new-current-uuid --clear-bitmap r1/0
sudo mount /precious
```

#### NFS

`./setup.sh setup_nfs`

On the node that has the DRBD volume `/precious` mounted, set the IP of the NFS server to be used by k8s:

```sh
sudo ip addr add 10.53.101.100/24 dev enp5s0.4001
```

#### K8S

For the first node `./setup.sh setup_k8s`. For nodes joining the cluster `./setup.sh setup_k8s 6` where `hetzner06` is an existing node.

- [metallb](https://metallb.universe.tf) instead of the default load balancer because it does not allow for a public IP different from the `k8s` node IP.
  `./setup.sh setup_k8s_metallb`
- [traefik](https://traefik.io/) requests with [annotations](https://github.com/traefik/traefik-helm-chart/blob/7a13fc8a61a6ad30fcec32eec497dab9d8aea686/traefik/values.yaml#L736) specific IPs from `metalldb`.
  `./setup.sh setup_k8s_traefik`
- [cert-manager](https://cert-manager.io/).
  `./setup.sh setup_k8s_certmanager`
- NFS storage class
  `./setup.sh setup_k8s_nfs`

#### Forgejo

[forgejo](https://code.forgejo.org/forgejo-helm/forgejo-helm) configuration in [ingress](https://code.forgejo.org/forgejo-helm/forgejo-helm#ingress) for the reverse proxy (`traefik`) to route the domain and for the ACME issuer (`cert-manager`) to obtain a certificate. And in [service](https://code.forgejo.org/forgejo-helm/forgejo-helm#service) for the `ssh` port to be bound to the desired IPs of the load balancer (`metallb`).

```
ingress:
  enabled: true
  annotations:
    # https://cert-manager.io/docs/usage/ingress/#supported-annotations
    # https://github.com/cert-manager/cert-manager/issues/2239
    cert-manager.io/cluster-issuer: letsencrypt-http
    cert-manager.io/private-key-algorithm: ECDSA
    cert-manager.io/private-key-size: 384
    kubernetes.io/ingress.class: traefik
    traefik.ingress.kubernetes.io/router.entrypoints: websecure
  tls:
    - hosts:
        - t1.forgejo.org
      secretName: tls-forgejo-t1-ingress-http
  hosts:
    - host: t1.forgejo.org
      paths:
        - path: /
          pathType: Prefix

service:
  http:
    type: ClusterIP
    ipFamilyPolicy: PreferDualStack
    port: 3000
  ssh:
    type: LoadBalancer
    annotations:
      metallb.universe.tf/loadBalancerIPs: 188.40.16.47,2a01:4f8:fff2:48::2
      metallb.universe.tf/allow-shared-ip: "key-to-share-failover"
    ipFamilyPolicy: PreferDualStack
    port: 2222
```

### K8S NFS storage creation

Define the 20GB `forgejo-data` pvc owned by user id 1000.

```sh
./setup.sh setup_k8s_pvc forgejo-data 20Gi 1000
```

[Instruct the forgejo pod](https://code.forgejo.org/forgejo-helm/forgejo-helm#persistence) to use the `forgejo-data` pvc.

```yaml
persistence:
  enabled: true
  create: false
  claimName: forgejo-data
```

## Disaster recovery and maintenance

### When a machine or disk is scheduled for replacement.

* `kubectl drain hetzner05` # evacuate all the pods out of the node to be shutdown
* `kubectl taint nodes hetzner05 key1=value1:NoSchedule` # prevent any pod from being created there (metallb speaker won't be drained, for instance)
* `kubectl delete node hetzner05` # let the cluster know it no longer exists so a new one by the same name can replace it

### Routing the failover IP

When the machine to which the failover IP (failover.forgejo.org) is routed is unavailable or to be shutdown, to the [Hetzner server panel](https://robot.hetzner.com/server), to the IPs tab and change the route of the failover IP to another node. All nodes are configured with the failover IP, there is nothing else to do.

### Manual boot operations

#### On the machine that runs the NFS server

* `sudo drbdadm primary r1` # Switch the DRBD to primary
* `sudo mount /precious` # DRBD volume shared via NFS
* `sudo ip addr add 10.53.101.100/24 dev enp5s0.4001` # add NFS server IP

#### On the other machines

* `sudo ip addr del 10.53.101.100/24 dev enp5s0.4001` # remove NFS server IP