JupyterLab on Agent-Sandbox

Deploy single-user JupyterLab environments using Kubernetes agent-sandbox.

Prerequisites

You must have agent-sandbox already installed on your cluster. Follow the installation guide:

Make sure the agent-sandbox controller is running before proceeding:

kubectl get pods -n agent-sandbox-system
# Should show agent-sandbox-controller-0 in Running state

Project Structure

.
├── README.md                    # This file
├── ../../INSTALL.md             # Agent-sandbox installation guide
├── jupyterlab.yaml              # Modular deployment (Secret + Sandbox only)
├── jupyterlab-full.yaml         # All-in-one deployment (+ ConfigMap with file contents)
└── files/
    ├── download_models.py       # Script to download HuggingFace models
    ├── requirements.txt         # Python dependencies
    └── welcome.ipynb            # Sample notebook

Installation Methods

Choose one of two deployment methods:

Method A: All-in-One Deployment

Use jupyterlab-full.yaml which contains everything in a single file.

Steps:

Edit the HuggingFace token:

# Open jupyterlab-full.yaml and replace HF_TOKEN with your actual token
vi jupyterlab-full.yaml

Or use sed:

export HF_TOKEN="your_actual_HF_token_here"
sed -i "s/HF_TOKEN/$HF_TOKEN/g" jupyterlab-full.yaml

Deploy everything:
```
kubectl apply -f jupyterlab-full.yaml
```
Skip to “Verify Installation” section below.

Method B: Modular Deployment

Use jupyterlab.yaml + create ConfigMap from files. This keeps YAML clean.

Steps:

Create the ConfigMap from files:

kubectl create configmap jupyter-init-files \
  --from-file=files/download_models.py \
  --from-file=files/requirements.txt \
  --from-file=files/welcome.ipynb \
  --namespace=default

Verify the ConfigMap was created:

kubectl get configmap jupyter-init-files -o yaml

Edit the HuggingFace token:

# Open jupyterlab.yaml and replace HF_TOKEN with your actual token
vi jupyterlab.yaml

Or use sed:

export HF_TOKEN="your_actual_HF_token_here"
sed -i "s/HF_TOKEN/$HF_TOKEN/g" jupyterlab.yaml

Deploy the Sandbox:
```
kubectl apply -f jupyterlab.yaml
```

Verify Installation

1. Monitor Initialization

The init container will download packages and models on first run (~5-10 minutes):

# Watch pod creation
kubectl get pods -l sandbox=jupyterlab -w

# Check init container logs
kubectl logs -f jupyterlab-sandbox -c setup-environment

# You should see:
# "Installing Python dependencies to persistent storage..."
# "Downloading models to persistent storage..."
# "Initialization complete!"

2. Check Main Container

Once init completes, verify JupyterLab started:

# Check main container logs
kubectl logs -f jupyterlab-sandbox -c jupyterlab

# Look for: "Jupyter Server ... is running at http://0.0.0.0:8888"

3. Verify Pod is Running

kubectl get sandbox jupyterlab-sandbox
kubectl get pod jupyterlab-sandbox

# Both should show Running status

Access JupyterLab

Port Forward to Local Machine

kubectl port-forward --address 0.0.0.0 pod/jupyterlab-sandbox 8888:8888

Access in browser:

Local: http://localhost:8888

Navigate to the welcome notebook:

Open http://localhost:8888/lab/tree/work/welcome.ipynb
Or click welcome.ipynb in the JupyterLab file browser

No password required (authentication disabled for sandbox use).

What Gets Deployed

Resources

Secret (jupyter-hf-token): Stores your HuggingFace token
ConfigMap (jupyter-init-files): Contains Python scripts and notebooks
Sandbox (jupyterlab-sandbox): Creates the JupyterLab environment with:
- Init container: Downloads models and installs dependencies
- Main container: Runs JupyterLab server
- PVC: 20Gi persistent storage

Customization

Add More Python Packages

Edit files/requirements.txt:

transformers>=4.51.0
torch
huggingface_hub
ipywidgets
pandas>=2.0.0        # Add your packages
scikit-learn
matplotlib

Then recreate the ConfigMap and force re-initialization:

# Method A (all-in-one): Edit jupyterlab-full.yaml and reapply
kubectl apply -f jupyterlab-full.yaml

# Method B (modular): Recreate ConfigMap
kubectl delete configmap jupyter-init-files
kubectl create configmap jupyter-init-files \
  --from-file=files/download_models.py \
  --from-file=files/requirements.txt \
  --from-file=files/welcome.ipynb

# Force re-init for both methods
kubectl exec jupyterlab-sandbox -- rm /home/jovyan/.initialized
kubectl delete pod jupyterlab-sandbox

Add More Models

Edit files/download_models.py:

model_names = [
    "Qwen/Qwen3-Embedding-0.6B",
    "meta-llama/Llama-3.2-1B-Instruct",                          # Add more
]

Then recreate ConfigMap and force re-init (same steps as above).

Add More Notebooks

Add .ipynb files to the files/ directory, then:

# Recreate ConfigMap with all files
kubectl delete configmap jupyter-init-files
kubectl create configmap jupyter-init-files \
  --from-file=files/ \
  --namespace=default

# Update init script to copy all notebooks
# Edit jupyterlab.yaml or jupyterlab-full.yaml:
# Change: cp /config/welcome.ipynb /home/jovyan/work/welcome.ipynb
# To:     cp /config/*.ipynb /home/jovyan/work/

# Reapply and restart
kubectl apply -f jupyterlab.yaml
kubectl delete pod jupyterlab-sandbox

Change Storage Size

Modify volumeClaimTemplates in the Sandbox spec:

volumeClaimTemplates:
  - metadata:
      name: workspace
    spec:
      resources:
        requests:
          storage: 50Gi  # Increase for large datasets

Note: You must delete the existing Sandbox and PVC to change storage size:

kubectl delete sandbox jupyterlab-sandbox
kubectl delete pvc jupyterlab-sandbox-workspace
kubectl apply -f jupyterlab.yaml

Troubleshooting

Init Container Fails

Check init logs:

kubectl logs jupyterlab-sandbox -c setup-environment

Common issues:

Invalid HuggingFace token - 401 Unauthorized errors
- Solution: Verify your token at https://huggingface.co/settings/tokens
- Update the secret: kubectl delete secret jupyter-hf-token && kubectl create secret ...
Out of disk space - Ephemeral storage exceeded
- This shouldn’t happen with current config (everything uses PVC)
- Check PVC is mounted: kubectl describe pod jupyterlab-sandbox | grep Volumes
Model download timeout - Network issues or large model
- Check init logs for specific error
- Try with a smaller model first

JupyterLab Not Starting

Check main container logs:

kubectl logs jupyterlab-sandbox -c jupyterlab