Container Images

Kubeflow Notebooks natively supports three types of notebooks, JupyterLab, RStudio, and Visual Studio Code (code-server), but any web-based IDE should work. Notebook servers run as containers inside a Kubernetes Pod, which means the type of IDE (and which packages are installed) is determined by the Docker image you pick for your server.

Official Images

Kubeflow provides a number of example container images to get you started with Kubeflow Notebooks.

This chart shows how the images are related to each other (note, the nodes are clickable links to the Dockerfiles):

%%{init: {'theme':'forest'}}%%
graph TD
  Base[Base] --> Jupyter[Jupyter]
  Base --> Code-Server[code-server]
  Base --> RStudio[RStudio]
  
  Jupyter --> PyTorch[PyTorch]
  Jupyter --> SciPy[SciPy]
  Jupyter --> TensorFlow[TensorFlow]
  
  Code-Server --> Code-Server-Conda-Python[Conda Python]
  RStudio --> Tidyverse[Tidyverse]

  PyTorch --> PyTorchFull[PyTorch Full]
  TensorFlow --> TensorFlowFull[TensorFlow Full]

  Jupyter --> PyTorchCuda[PyTorch CUDA]
  Jupyter --> TensorFlowCuda[TensorFlow CUDA]

  PyTorchCuda --> PyTorchCudaFull[PyTorch CUDA Full]
  TensorFlowCuda --> TensorFlowCudaFull[TensorFlow CUDA Full]

  click Base "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/base"
  click Jupyter "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter"
  click Code-Server "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/codeserver"
  click RStudio "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/rstudio"
  click PyTorch "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-pytorch"
  click SciPy "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-scipy"
  click TensorFlow "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-tensorflow"
  click Code-Server-Conda-Python "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/codeserver-python"
  click Tidyverse "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/rstudio-tidyverse"
  click PyTorchFull "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-pytorch-full"
  click TensorFlowFull "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-tensorflow-full"
  click PyTorchCuda "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-pytorch-cuda"
  click TensorFlowCuda "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-tensorflow-cuda"
  click PyTorchCudaFull "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-pytorch-cuda-full"
  click TensorFlowCudaFull "https://github.com/kubeflow/kubeflow/tree/master/components/example-notebook-servers/jupyter-tensorflow-cuda-full"

Base Images

These images provide a common starting point for Kubeflow Notebook containers.

Kubeflow Images

These images extend the base images with common packages used in the real world.

Package Installation

Packages installed by users after spawning a Kubeflow Notebook will only last the lifetime of the pod (unless installed into a PVC-backed directory).

To ensure packages are preserved throughout Pod restarts users will need to either:

1. Build custom images that include them, or
2. Ensure they are installed in a PVC-backed directory

Custom Images

You can build your own custom images to use with Kubeflow Notebooks.

The easiest way to ensure your custom image meets the requirements is to extend one of our base images.

Image Requirements

For a container image to work with Kubeflow Notebooks, it must:

• expose an HTTP interface on port 8888:
  • kubeflow sets an environment variable NB_PREFIX at runtime with the URL path we expect the container be listening under
  • kubeflow uses IFrames, so ensure your application sets Access-Control-Allow-Origin: * in HTTP response headers
• run as a user called jovyan:
  • the home directory of jovyan should be /home/jovyan
  • the UID of jovyan should be 1000
• start successfully with an empty PVC mounted at /home/jovyan:
  • kubeflow mounts a PVC at /home/jovyan to keep state across Pod restarts

Install Python Packages

You may extend one of the images and install any pip or conda packages your Kubeflow Notebook users are likely to need. As a guide, look at ./jupyter-pytorch-full/Dockerfile for a pip install ... example, and the ./rstudio-tidyverse/Dockerfile for conda install ....

A common cause of errors is users running pip install --user ..., causing the home-directory (which is backed by a PVC) to contain a different or incompatible version of a package contained in /opt/conda/...

Install Linux Packages

You may extend one of the images and install any apt-get packages your Kubeflow Notebook users are likely to need. Ensure you swap to root in the Dockerfile before running apt-get, and swap back to $NB_USER after.

Configure S6 Overlay

Some use-cases might require custom scripts to run during the startup of the Notebook Server container, or advanced users might want to add additional services that run inside the container (for example, an Apache or NGINX web server). To make this easy, we use the s6-overlay.

The s6-overlay differs from other init systems like tini. While tini was created to handle a single process running in a container as PID 1, the s6-overlay is built to manage multiple processes and allows the creator of the image to determine which process failures should silently restart, and which should cause the container to exit.

Create Scripts

Scripts that need to run during the startup of the container can be placed in /etc/cont-init.d/, and are executed in ascending alphanumeric order.

An example of a startup script can be found in ./rstudio/s6/cont-init.d/02-rstudio-env-fix. This script uses the with-contenv helper so that environment variables (passed to container) are available in the script. The purpose of this script is to snapshot any KUBERNETES_* environment variables into the Renviron.site at pod startup, as without these variables kubectl does not work.

Create Services

Extra services to be monitored by s6-overlay should be placed in their own folder under /etc/services.d/ containing a script called run and optionally a finishing script finish.

An example of a service can be found in the run script of .jupyter/s6/services.d/jupyterlab which is used to start JupyterLab itself. For more information about the run and finish scripts, please see the s6-overlay documentation.

Run Services As Root

There may be cases when you need to run a service as root, to do this, you can change the Dockerfile to have USER root at the end, and then use s6-setuidgid to run the user-facing services as $NB_USER.

Our example images run s6-overlay as $NB_USER (not root), meaning any files or scripts related to s6-overlay must be owned by the $NB_USER user to successfully run.

Next steps

• Use your container image by specifying it when spawning your notebook server. (See the quickstart guide.)