K3d

k3d - Kubernetes Clusters

Overview

The k3d command manages lightweight Kubernetes (k3s) clusters running in Podman containers. Clusters are joined to the bazzite-ai network, enabling DNS-based service discovery with other pods (ollama, jupyter, etc.).

Key Concept: k3d wraps k3s (lightweight Kubernetes) in containers, providing fast cluster creation with full Kubernetes API compatibility. GPU passthrough is supported via NVIDIA device plugin.

Quick Reference

Parameters

Configuration

# Default: Single-node cluster, port 6443, auto-detect GPU
ujust k3d config

# Custom API port (long form)
ujust k3d config --port=6444

# Custom API port (short form)
ujust k3d config -p 6444

# Add 2 agent (worker) nodes
ujust k3d config --agents=2

# With GPU support (NVIDIA)
ujust k3d config --gpu-type=nvidia

# Combine parameters (short form)
ujust k3d config -p 6444 -a 2 -g nvidia

# Network-wide API access
ujust k3d config --bind=0.0.0.0

# Custom ingress ports
ujust k3d config --http-port=8080 --https-port=8443

# Disable built-in components
ujust k3d config --disable-traefik --disable-servicelb

# Specific k3s version
ujust k3d config --k3s-version=v1.28.5+k3s1

Update Existing Configuration

Running config when already configured will recreate the cluster with new settings. Data is not preserved - export important resources first.

Lifecycle Commands

Start/Stop/Restart

# Start cluster (instance 1 default)
ujust k3d start

# Start specific instance (long form)
ujust k3d start --instance=1

# Start specific instance (short form)
ujust k3d start -n 1

# Stop cluster
ujust k3d stop

# Restart cluster
ujust k3d restart

View Logs

# View k3s server logs (default 50 lines)
ujust k3d logs

# More lines (long form)
ujust k3d logs --lines=100

# More lines (short form)
ujust k3d logs -l 100

# Specific instance
ujust k3d logs -n 2 -l 100

Status

# Show cluster status and node list
ujust k3d status

# List all clusters
ujust k3d list

Shell Access (kubectl)

Execute kubectl commands directly in the cluster context:

# Interactive shell with kubectl available
ujust k3d shell

# Run kubectl commands (use -- separator)
ujust k3d shell -- kubectl get nodes
ujust k3d shell -- kubectl get pods -A
ujust k3d shell -- kubectl apply -f deployment.yaml

# Specific instance
ujust k3d shell -n 2 -- kubectl get services

GPU Support

k3d supports GPU passthrough for NVIDIA GPUs via the NVIDIA device plugin.

Setup GPU

# Install NVIDIA device plugin in cluster
ujust k3d gpu

# Specific instance
ujust k3d gpu --instance=2

Verify GPU Access

# Check device plugin is running
ujust k3d shell -- kubectl get pods -n kube-system | grep nvidia

# Run GPU test pod
ujust k3d shell -- kubectl run gpu-test --image=nvidia/cuda:12.0-base \
  --restart=Never --rm -it --command -- nvidia-smi

GPU Pod Example

apiVersion: v1
kind: Pod
metadata:
  name: gpu-pod
spec:
  containers:
  - name: cuda-container
    image: nvidia/cuda:12.0-base
    resources:
      limits:
        nvidia.com/gpu: 1

bazzite-ai Network Integration

k3d clusters join the bazzite-ai network, enabling service discovery:

# From inside k8s pods, access other bazzite-ai services:
curl http://ollama:11434/api/tags
curl http://jupyter:8888/

Available DNS names:

• ollama:11434 - Ollama API
• jupyter:8888 - JupyterLab
• openwebui:3000 - Open WebUI
• comfyui:8188 - ComfyUI

Multi-Instance

Run multiple k3d clusters simultaneously:

# Create first cluster
ujust k3d config -n 1 -p 6443

# Create second cluster with different port
ujust k3d config -n 2 -p 6444

# List all clusters
ujust k3d list

# Interact with specific cluster
ujust k3d shell -n 2 -- kubectl get nodes

Built-in Components

k3d includes these components by default:

Common Workflows

Initial Setup

# 1. Create cluster
ujust k3d config

# 2. Start cluster
ujust k3d start

# 3. Verify nodes
ujust k3d shell -- kubectl get nodes

# 4. Deploy application
ujust k3d shell -- kubectl apply -f my-app.yaml

GPU ML Workloads

# 1. Create cluster with GPU
ujust k3d config --gpu-type=nvidia

# 2. Start and setup GPU support
ujust k3d start
ujust k3d gpu

# 3. Deploy GPU workload
ujust k3d shell -- kubectl apply -f gpu-deployment.yaml

Development Cluster

# Fast local cluster for testing
ujust k3d config -a 1  # 1 server + 1 agent

# Deploy test app
ujust k3d shell -- kubectl create deployment nginx --image=nginx
ujust k3d shell -- kubectl expose deployment nginx --port=80 --type=LoadBalancer

# Access via traefik
curl http://localhost

Troubleshooting

Cluster Won't Start

Check:

ujust k3d status
ujust k3d logs --lines=100

Common causes:

• Port already in use
• Podman not running
• Insufficient resources

Fix:

# Use different port
ujust k3d delete
ujust k3d config --port=6444
ujust k3d start

GPU Not Available in Pods

Symptom: Pods requesting GPU resources stay Pending

Check:

ujust k3d shell -- kubectl describe pod <pod-name>
ujust k3d shell -- kubectl get pods -n kube-system | grep nvidia

Fix:

# Reinstall device plugin
ujust k3d gpu

# Or recreate cluster with GPU
ujust k3d delete
ujust k3d config --gpu-type=nvidia
ujust k3d start
ujust k3d gpu

Service Discovery Fails

Symptom: Can't reach ollama:11434 from k8s pods

Check:

# Verify network membership
podman network inspect bazzite-ai

Fix:

# Recreate cluster (joins network)
ujust k3d delete
ujust k3d config
ujust k3d start

kubectl Connection Refused

Symptom: kubectl commands fail with "connection refused"

Check:

ujust k3d status

Fix:

# Restart cluster
ujust k3d restart

Configuration Files

Cross-References

• Related Skills: portainer (container UI), ollama (LLM inference)
• k3d Docs: https://k3d.io/
• k3s Docs: https://docs.k3s.io/
• GPU Setup: ujust config gpu setup

When to Use This Skill

Use when the user asks about:

• "kubernetes", "k8s", "k3d", "k3s"
• "local kubernetes", "dev cluster"
• "kubectl", "deploy to kubernetes"
• "kubernetes GPU", "GPU pods"
• "start k3d", "configure k3d", "k3d not working"