Custom Jupyter Image

Create, publish, and use your own Docker image for Kubeflow.

---

• Custom packages

  ---

  Customize environment to your needs.

• Reproducibility

  ---

  Same environment for the entire team.

• Teaching

  ---

  Pre-prepared images for courses.

• Research

  ---

  Specific ML libraries and tools.

---

Dockerfile

Base image

# NOTE: Use Makefile or buildx for proper build steps.
ARG BASE_IMG=ghcr.io/kubeflow/kubeflow/notebook-servers/jupyter-pytorch:latest
FROM ${BASE_IMG}

USER root
ENV DEBIAN_FRONTEND=noninteractive

# 1) System tools and libraries
RUN apt-get update \
 && apt-get install -y --no-install-recommends \
    sudo git curl wget vim nano htop unzip \
    build-essential cmake pkg-config \
    libssl-dev libffi-dev libxml2-dev libxslt1-dev zlib1g-dev \
    libjpeg-dev libpng-dev libtiff-dev \
    libblas-dev liblapack-dev gfortran python3-dev \
    libgl1 libglib2.0-0 ffmpeg graphviz \
 && rm -rf /var/lib/apt/lists/*

# 2) Passwordless sudo for NB_USER
RUN usermod -aG sudo "${NB_USER}" \
 && echo "${NB_USER} ALL=(ALL) NOPASSWD:ALL" > /etc/sudoers.d/99-nbuser \
 && chmod 0440 /etc/sudoers.d/99-nbuser

# 3) Basic mamba/conda packages
RUN mamba install -y -q --freeze-installed \
    numpy scipy pandas scikit-learn scikit-image seaborn \
    sympy statsmodels networkx \
 && mamba clean -a -f -y

# 4) Pip packages (ML/DL)
ENV PIP_NO_CACHE_DIR=1
ENV MPLBACKEND=Agg
RUN python -m pip install -U pip setuptools wheel \
 && python -m pip install \
    matplotlib ipympl bokeh plotly \
    opencv-python-headless \
    xgboost lightgbm \
    nltk spacy \
    transformers datasets \
    pytorch-lightning

# 5) Custom requirements
COPY --chown=${NB_USER}:${NB_GID} requirements.txt /tmp/requirements.txt
RUN python -m pip install -r /tmp/requirements.txt \
 && rm -f /tmp/requirements.txt

# 6) Back to NB_USER
USER ${NB_USER}

---

requirements.txt

# Kubeflow Pipelines SDK
kfp==2.11.0

# JupyterLab extensions
jupyterlab-git==0.50.2

---

Build and Push

Makefile

IMAGE ?= repository.cloud.tuke.sk/ml/jupyter-pytorch-plus
TAG   ?= v1.0.0

.PHONY: build push

build:
    docker build -t $(IMAGE):$(TAG) .

push:
    docker push $(IMAGE):$(TAG)

Commands

# Build
make build IMAGE=repository.cloud.tuke.sk/ml/jupyter-pytorch-plus TAG=v1.0.0

# Login to registry
docker login repository.cloud.tuke.sk

# Push
make push IMAGE=repository.cloud.tuke.sk/ml/jupyter-pytorch-plus TAG=v1.0.0

---

Usage in Kubeflow

Create new notebook

In Kubeflow UI, go to Notebooks → New Notebook.

Enter Custom Image

In the Custom Image field, enter:

repository.cloud.tuke.sk/ml/jupyter-pytorch-plus:v1.0.0

Set resources and launch

Set CPU, RAM, and Workspace Volume, then click Launch.

---

What the Image Contains

System libraries (APT)

Package	Purpose
build-essential, cmake	Python package compilation
libjpeg, libpng, libtiff	Image processing
ffmpeg	Audio/video, OpenCV
graphviz	Graphs, DAG visualizations

Conda/Mamba packages

Package	Purpose
numpy, scipy, pandas	Numerical computing
scikit-learn, scikit-image	Machine learning
seaborn, sympy, statsmodels	Visualization, statistics
networkx	Graph algorithms

Pip packages (ML/DL)

Package	Purpose
matplotlib, bokeh, plotly	Visualization
opencv-python-headless	Computer vision
xgboost, lightgbm	Gradient boosting
nltk, spacy	NLP
transformers, datasets	Hugging Face ecosystem
pytorch-lightning	PyTorch wrapper

---

Practical Recommendations

Best practices

• Pin versions – especially PyTorch, CUDA, HF transformers
• Don't download models in build phase – they increase image size
• Use CHANGELOG and versions like v1.0.0
• Don't use :latest in production
• Put frequently changing things lower in Dockerfile

---

Post-deploy Check

python -c "import torch, transformers, sklearn, cv2; \
  print('Torch:', torch.__version__); \
  print('Transformers:', transformers.__version__); \
  print('cv2:', cv2.__version__)"
python -c "import kfp; print('KFP:', kfp.__version__)"
jupyter labextension list

---

Code-server Image (VS Code)

For VS Code in browser, create a similar image with code-server:

IMAGE ?= repository.cloud.tuke.sk/dev/codeserver
TAG   ?= v1.0.0

build:
    docker build -t $(IMAGE):$(TAG) .

push:
    docker push $(IMAGE):$(TAG)

Local test

docker run --rm -p 8888:8888 \
  -e PASSWORD='strong-password' \
  repository.cloud.tuke.sk/dev/codeserver:v1.0.0 \
  code-server --bind-addr 0.0.0.0:8888 --auth password /home/jovyan

---

Next Steps

• Creating notebook

  ---

  Launch notebook with custom image.

  Guide

• FAQ

  ---

  Common questions about ML TUKE.

  FAQ