k3s

name: k3s description: Use when installing or managing a self-hosted Kubernetes cluster with k3s, setting up kubectl or kubeconfig, adding server or agent nodes, or choosing between single-node and multi-node topologies. Also trigger on embedded etcd vs SQLite datastore decisions, k3s systemd service configuration, or accessing a cluster remotely. NOT for workload resources like Deployments or Services (see k8s-workloads), networking and ingress (see k8s-networking), storage (see k8s-storage), Helm (see helm), or containerizing applications (see docker).

k3s

Patterns and decisions for installing, configuring, and scaling k3s Kubernetes clusters in self-hosted environments.

Distribution Decision

Choose k3s when self-hosting on VPS, homelab, Raspberry Pi, or edge — single binary (~65 MB), reasonable defaults, supports single-node through large clusters, CNCF-maintained, production-ready.

Alternatives and when to prefer them:

• minikube / Docker Desktop / Orbstack embedded K8s — local dev/learning only; not for production
• k0s — similar philosophy to k3s, less mature ecosystem
• kubeadm — full control over every component; only if you have specific requirements and operational expertise
• Cloud-managed (EKS, GKE, AKS) — when you don't want to manage the control plane; costs more but eliminates operational overhead
• Docker Compose / Kamal — single-host only, no self-healing or rolling updates; fine for simple apps that don't need cluster features

k3s defaults you should know:

• Container runtime: containerd
• CNI: Flannel (VXLAN)
• DNS: CoreDNS
• Ingress: Traefik (disabled in this guide — use ingress-nginx instead)
• Datastore: SQLite (single-node) or embedded etcd (HA, requires --cluster-init)
• Storage: Local Path Provisioner

Core Concepts (Brief)

Control plane components (run on server nodes): etcd, kube-apiserver (port 6443), kube-scheduler, kube-controller-manager

Worker components (run on all nodes): kubelet, containerd, kube-proxy

In k3s, server nodes run both by default. Control-plane-only nodes require taints or the experimental --disable-agent flag.

Installation: Single-Node

# On the server (Ubuntu 24.04, as root)
apt install curl -y
# Pin the version — check https://github.com/k3s-io/k3s/releases for the latest stable
# Replace YOUR_PUBLIC_IP with the server's public IP
curl -sfL https://get.k3s.io | INSTALL_K3S_VERSION=v1.32.3+k3s1 sh -s - server \
  --disable=traefik \
  --cluster-init \
  --cluster-domain=cluster.local \
  --tls-san=YOUR_PUBLIC_IP

# Verify
k3s kubectl get nodes
# NAME              STATUS   ROLES                       AGE   VERSION
# node0.example.com Ready    control-plane,etcd,master   96s   v1.32.3+k3s1

Flag decisions:

• INSTALL_K3S_VERSION — always pin to a specific release; without it, the install script pulls the latest version, which may differ from what you tested or from what other nodes in your cluster are running; version skew between nodes causes instability
• --disable=traefik — remove the default ingress controller; install ingress-nginx separately for more control
• --cluster-init — use embedded etcd instead of SQLite; required if you ever want to add more control-plane nodes; enables automatic migration path
• --cluster-domain — set a custom cluster domain for CoreDNS; must match across all nodes; defaults to cluster.local
• --tls-san — add your server's public IP/hostname to the TLS certificate SAN list; required for remote kubectl access; without it, kubectl fails with an x509 certificate error when connecting via the public IP

If installation fails: Run /usr/local/bin/k3s-uninstall.sh before retrying to ensure a clean state. The uninstall script removes all k3s data including etcd.

Datastore choice:

• SQLite (default, no --cluster-init) — simplest; fine for single-node that will never scale to HA; single file, easy to back up
• Embedded etcd (--cluster-init) — required for multi-server HA; adds ~100 MB RAM overhead; use this unless resources are extremely constrained
• External DB (--datastore-endpoint) — use if your team already manages PostgreSQL/MySQL; reduces control-plane node resource requirements

Static IP required. Kubernetes needs a stable node IP. On DHCP networks, assign a static IP before installing.

kubeconfig Setup

k3s auto-generates /etc/rancher/k3s/k3s.yaml on install. The server address defaults to 127.0.0.1:6443 — update it for remote access.

Remote access (recommended pattern):

# On the server: print the kubeconfig
cat /etc/rancher/k3s/k3s.yaml

Copy the output to your local machine, save as e.g. ~/clusters/mycluster.yaml. Then open the file and change the .clusters[0].cluster.server value from https://127.0.0.1:6443 to your server's actual public IP or hostname:

# Before:
server: https://127.0.0.1:6443
# After:
server: https://YOUR_PUBLIC_IP:6443

# On local machine: export before each session
export KUBECONFIG=/Users/you/clusters/mycluster.yaml
kubectl get nodes

Three ways to supply kubeconfig (in order of preference for self-hosting):

1. export KUBECONFIG=<absolute-path> — most flexible; works with kubectl, helm, k9s, kubectx; set at the start of each shell session
2. --kubeconfig <path> flag — per-command; impractical for regular use
3. ~/.kube/config — convenient for single-cluster local dev; not recommended when managing multiple clusters

Store kubeconfig outside the project directory or add it to .gitignore and .dockerignore. It contains credentials and certificates.

Certificate expiry: The user cert expires after 365 days and is renewed automatically the next time k3s restarts (within 90 days of expiry). On long-running clusters, restart k3s before the cert expires: systemctl restart k3s. After renewal, re-copy the kubeconfig file. Check cert status: k3s certificate check --output table

If the k3s service fails to start after cert expiry (cert expired while service was stopped): run k3s certificate rotate as root, then systemctl restart k3s.

kubectl Essentials

# Cluster state
kubectl get nodes                          # list nodes with roles and status
kubectl get nodes --show-labels            # include node labels
kubectl cluster-info                       # control plane and DNS addresses
kubectl version                            # client and server versions

# Resources (works for any resource type)
kubectl get pods                           # pods in current namespace
kubectl get pods -n kube-system            # specific namespace
kubectl get pods -A                        # all namespaces
kubectl get pods -A -w                     # watch for changes
kubectl describe pod <name>                # detailed state + events
kubectl logs <pod> -f                      # stream logs
kubectl logs <pod> -c <container>          # specific container in multi-container pod

# Namespaces
kubectl get namespaces                     # list all
kubectl create namespace db
kubectl config set-context --current --namespace=db   # switch default namespace
kubectl config view --minify | grep namespace:        # confirm current namespace

# Apply / delete
kubectl apply -f manifest.yaml             # create or update
kubectl delete -f manifest.yaml            # delete resources defined in file
kubectl delete pod <name>                  # delete a specific resource

# Debugging
kubectl exec -it <pod> -- /bin/sh          # shell into a pod
kubectl port-forward pod/<name> 8080:80    # forward local port to pod
kubectl get events -A -w                   # watch all events (essential for troubleshooting)

# Context management
kubectl config get-contexts                # list all contexts
kubectl config use-context <name>          # switch context
kubectl config set-context --current --namespace=<ns>  # change namespace in current context

Namespace-scoped vs cluster-scoped resources:

• Namespace-scoped: Pods, Deployments, Services, ConfigMaps, Secrets — always specify -n or -A if not in the right namespace
• Cluster-scoped: Nodes, PersistentVolumes, StorageClasses, Namespaces — no -n flag needed

# List all cluster-scoped resources
kubectl api-resources --namespaced=false

For a full kubectl cheatsheet, see references/kubectl-cheatsheet.md.

Namespace Conventions

Four built-in namespaces: default (educational only), kube-system (system components), kube-node-lease (heartbeats), kube-public (public certs).

Recommended starting namespaces for small clusters:

kubectl create namespace db          # databases (PostgreSQL, Redis, etc.)
kubectl create namespace production  # production workloads
kubectl create namespace test        # test/staging workloads
kubectl create namespace misc        # anything that doesn't fit elsewhere

Never deploy production workloads to default. It has no RBAC boundaries and is a common source of accidental deletions.

Single-Node vs Multi-Node Decision

Single node (server only)
├── Pros: simple, cheap, easy to manage
├── Cons: no HA — control plane loss = no cluster management (pods keep running)
└── Use when: dev/staging, low-traffic production, budget-constrained

2 nodes (1 server + 1 agent)
├── Pros: workload isolation from control plane
├── Cons: 1 etcd member, no fault tolerance — losing the server destroys cluster state
└── Avoid: worse than 1 or 3; not recommended

3 server nodes (recommended HA minimum)
├── Pros: quorum = 2, survives 1 node loss; full HA
├── Cons: 3x control-plane resource overhead
└── Use when: production requiring HA

3 server + N agent nodes
├── Pros: scales workload capacity without affecting quorum
└── Use when: workload demands exceed 3 server nodes' capacity

Key rule: Always use an odd number of server nodes (1, 3, 5, 7...). Even numbers don't improve fault tolerance and risk split-brain.

For the full clustering guide (adding server and agent nodes step-by-step), see references/clustering.md.

k3s Service Management

# Service control
systemctl status k3s            # check status
systemctl restart k3s           # restart (on server nodes)
systemctl restart k3s-agent     # restart (on agent nodes)

# View startup flags
cat /etc/systemd/system/k3s.service

# Logs
journalctl -u k3s -f            # stream k3s logs
journalctl -u k3s-agent -f      # stream agent logs

# Token (needed to add nodes)
cat /var/lib/rancher/k3s/server/token

# Built-in tools
k3s kubectl get nodes           # kubectl without separate kubeconfig
k3s etcd-snapshot               # backup etcd state
k3s token create                # generate a bootstrap token (expires 24h)
k3s certificate check --output table  # check cert expiry

Required Ports (Firewall Rules)

Open these between nodes before adding them to the cluster. Missing ports are the most common cause of nodes stuck in NotReady.

Security: Port 6443 must be accessible between cluster nodes and your workstation, but must not be open to the public internet (0.0.0.0/0). Restrict it to known source IPs in your cloud firewall. For remote access from dynamic IPs, use a VPN or SSH tunnel (ssh -L 6443:localhost:6443 user@server) rather than exposing the API server directly.