Containers for Kubeflow

Containers to be used with Kubeflow for Data Science.

Introduction

Our Container images are based on the community driven jupyter/docker-stacks. This enables us to focus only on the additional toolsets that we require to enable our data scientists.

Usage

Generating Dockerfiles

Use make generate-dockerfiles to generate all Dockerfiles. These will be written to ./output/imagename, along with any required files for the build context

Building and Tagging Docker Images

Use make build/IMAGENAME to build an already generated (see above) Dockerfile. This by default generates images with:

repo=k8scc01covidacr.azurecr.io
tag=BRANCH_NAME For example: k8scc01covidacr.azurecr.io/IMAGENAME:BRANCH_NAME.

make build also accepts arguments for REPO and TAG to override these behaviours. For example, make build/jupyterlab-cpu REPO=myrepo TAG=notLatest.

make post-build/IMAGENAME is meant for anything that is commonly done after building an image, but currently only adds common tags. It adds tags of SHA, SHORT_SHA, and BRANCH_NAME to the given image, and accepts a SOURCE_FULL_IMAGE_NAME argument if you're trying to tag an existing image that has a non-typical name. For example:

make post-build/IMAGENAME will apply SHA, SHORT_SHA, and BRANCH_NAME tags to k8scc01covidacr.azurecr.io/IMAGENAME:BRANCH_NAME (eg: using the default REPO and TAG names)
make post-build/IMAGENAME SOURCE_FULL_IMAGE_NAME=oldRepo/oldImage:oldTag REPO=newRepo will make the following new aliases for oldRepo/oldImage:oldTag REPO=newRepo:
- newRepo/IMAGENAME:SHA
- newRepo/IMAGENAME:SHORT_SHA
- newRepo/IMAGENAME:BRANCH_NAME

Pulling and Pushing Docker Images

make pull/IMAGENAME and make push/IMAGENAME work similarly to make build/IMAGENAME. REPO and TAG arguments are available to override their default values.

Note: To use make pull or make push, you must first log in to ACR (az acr login -n k8scc01covidacr) Note: make push by default does docker push --all-tags in order to push the SHA, SHORT_SHA, etc., tags.

Testing images

Running and Connecting to Images Locally/Interactively

To test an image interactively, use make dev/IMAGENAME. This docker run's a built image, automatically forwarding ports to your local machine and providing a link to connect to.

Automated Testing

Automated tests are included for the generated Docker images using pytest. This testing suite is modified from the docker-stacks test suite. Image testing is invoked through make test/IMAGENAME (with optional REPO and TAG arguments like make build).

Testing of a given image consists of general and image-specific tests:

└── tests
    ├── general                             # General tests applied to all images
    │   └── some_general_test.py
    ├── jupyterlab-cpu                      # Test applied to a specific image
    │   └── some_jupyterlab-cpu-specific_test.py
    └── jupyterlab-tensorflow

Where tests/general tests are applied to all images, and tests/IMAGENAME are applied only to a specific image. Pytest will start the image locally and then run the provided tests to determine if Jupyterlab is running, python packages are working properly, etc. Tests are formatted using typical pytest formats (python files with def test_SOMETHING() functions). conftest.py defines some standard scaffolding for image management, etc.

General Development Workflow

Modifying Dockerfiles (local testing)

Clone the repo
(optional) make pull/IMAGENAME TAG=SOMEEXISTINGTAG to pull an existing version of the image you are working on (this could be useful as a build cache to reduce development time below)
Change an image via the docker-bits that are used to create it, not the files in the output/ folder. Same goes for the shell scripts and json files - they should be modified from the resources folder.
- For quick-iteration debugging you can directly edit the ./output files, but make sure you commit any changes you want to keep back to the ./docker-bits
After making your changes, generate new Dockerfiles through make generate-dockerfiles
Build your edited image using make build/IMAGENAME (or, if you pulled a version of it above, you can use make build/IMAGENAME DARGS="--cache-from SOMEOLDREPO/SOMEOLDIMAGE:SOMETAG", which will use layers from the pulled image as cached layers if possible, speeding up your build)
Test your image:
- using automated tests through make test/IMAGENAME
- manually by docker run --it -p 8888:8888 REPO/IMAGENAME:TAG, then opening it in http://localhost:8888

Modifying Dockerfiles (on-platform testing)

GitHub Actions CI is enabled to do building, scanning, automated testing, and (optionally) pushing of our images to ACR. Build, test, and scan CI triggers on:

any push to master
any push to an open PR This allows for easy scanning and automated testing for images.

GitHub Actions CI also enables pushing built images to our ACR, making them accessible from the platform. This occurs on:

any push to master
any push to an open PR that also has the auto-deploy label on the PR This allows developers to opt-in to on-platform testing. For example, when you need to build in github and test on platform (or want someone else to be able to pull your image):
open a PR and add the auto-deploy label
push to your PR and watch the GitHub Action CI
access your image in Kubeflow via a custom image from any of:
- k8scc01covidacr.azurecr.io/IMAGENAME:SHA
- k8scc01covidacr.azurecr.io/IMAGENAME:SHORT_SHA
- k8scc01covidacr.azurecr.io/IMAGENAME:BRANCH_NAME

Adding new Images

Dockerfiles are defined using make with recipes defined in the Makefile. They pull segments of Dockerfiles from docker-bits and assemble them. All output images should meet the following criteria:

be generated by calling make generate-dockerfiles
have outputs written to output/imagename, where imagename is a valid Docker image name (eg: all lowercase, no special characters)

To add new images, edit the makefile such that it generates the ./output/imagename directory. You can usually follow the existing recipes (or even add an extra piece to them), or you can add a whole new make target (but make sure to add your new target to make generate-dockerfiles as well).

Modifying and Testing CI

If making changes to CI that cannot be done on a branch (eg: changes to issue_comment triggers), you can:

fork the 'kubeflow-containers' repo

Modify the CI with

REGISTRY: (your own dockerhub repo, eg: "j-smith" (no need for the full url))

Change

- uses: azure/docker-login@v1
  with:
    login-server: ${{ env.REGISTRY_NAME }}.azurecr.io
    username: ${{ secrets.REGISTRY_USERNAME }}
    password: ${{ secrets.REGISTRY_PASSWORD }}

- uses: docker/login-action@v1
  with:
    username: ${{ secrets.REGISTRY_USERNAME }}
    password: ${{ secrets.REGISTRY_PASSWORD }}

In your forked repo, define secrets for REGISTRY_USERNAME and REGISTRY_PASSWORD with your dockerhub credentials (you should use an API token, not your actual dockerhub password)

Note: Since pushing comes right at the end of the CI, in many cases you don't need to have a valid registry to test the CI on a fork. It will fail on the push step, but all other steps will clearly work and you can know it should safely merge back into the main repo.

Other Development Notes

Set User File Permissions after Every `pip`/`conda` Install or Edit of User's Home Files

The Dockerfiles in this repo are intended to construct compute environments for a non-root user jovyan to ensure the end user has the least privileges required for their task, but installation of some of the software needed by the user must be done as the root user. This means that installation of anything that should be user editable (eg: pip and conda installs, additional files in /home/$NB_USER, etc.) will by default be owned by root and not modifiable by jovyan. Therefore we must change the permissions of these files to allow the user specific access for modification. For example, most pip install/conda install commands occur as the root user and result in new files in the $CONDA_DIR directory that will be owned by root and cause issues if user jovyan tried to update or uninstall these packages (as they by default will not have permission to change/remove these files).

To fix this issue, end any RUN command that edits any user-editable files with:

fix-permissions $CONDA_DIR && \
fix-permissions /home/$NB_USER

This fix edits the permissions of files in these locations to allow user access. Note that if these are not applied in the same layer as when the new files were added it will result in a duplication of data in the layer because the act of changing permissions on a file from a previous layer requires a copy of that file into the current layer. So something like:

RUN add_1GB_file_with_wrong_permissions_to_NB_USER.sh && \
	fix-permissions /home/$NB_USER

would add a single layer of about 1GB, whereas

RUN add_1GB_file_with_wrong_permissions_to_NB_USER.sh

RUN fix-permissions /home/$NB_USER

would add two layers, each about 1GB (2GB total).

Structure

.
├── Makefile                                # Cats the docker-bits together
│
├── docker-bits                             # The docker snippets. Numbering indicates the DAG.
│   ├── 0_CPU.Dockerfile
│   ├── 1_CUDA-11.0.Dockerfile
│   ├── 1_CUDA-11.1.Dockerfile
│   ├── 2_PyTorch.Dockerfile
│   ├── 2_Tensorflow.Dockerfile
│   ├── 3_Kubeflow.Dockerfile
│   ├── 4_CLI.Dockerfile
│   ├── 5_DB-Drivers.Dockerfile
│   ├── 6_JupyterLab.Dockerfile
│   ├── 6_RStudio.Dockerfile
│   ├── 6_JupyterLab-OL-compliant.Dockerfile
│   ├── 6_RemoteDesktop.Dockerfile
│   ├── ∞_CMD.Dockerfile
│   └── ∞_CMD_RemoteDesktop.Dockerfile
│
├── resources                               # the Docker context (files for COPY)
├── ├── common                              # files required by all images
│      ├── clean-layer.sh
│      ├── helpers.zsh
│      ├── jupyterlab-overrides.json
│      ├── landing_page
│      ├── nginx
│      ├── README.md
│      └── start-custom.sh
├── ├── remote-desktop                      # directory containing files only for the remote desktop
|      ├── desktop-files                    # desktop configuration 
|      ├── French                           # files to support i18n of remote desktop
|      ├── qgis-2020.gpg.key
|      └── start-remote-desktop.sh
|      
│
├── scripts                                 # Helper Scripts (NOT automated.)
├── ├── remote-desktop                      # Scripts installing applications on remote desktop
|      ├── firefox.sh
|      ├── fix-permissions.sh
|      ├── qgis.sh
|      ├── r-studio-desktop.sh
|      └── vs-code-desktop.sh
│   ├── CHECKSUMS
│   ├── checksums.sh
│   ├── get-nvidia-stuff.sh
│   ├── start-custom-OL-compliant.sh
│   └── README.md
│
└── output                                  # Staging area for a `docker build .`
    ├── JupyterLab-CPU/
    ├── JupyterLab-PyTorch/
    ├── JupyterLab-Tensorflow/
    |── RStudio/
    |── RemoteDesktop/
    ├── JupyterLab-CPU-OL-compliant/        # These images use JupyterLab 3.0 and contain only OL-compliant extensions
    ├── JupyterLab-PyTorch-OL-compliant/
    └── JupyterLab-Tensorflow-OL-compliant/
└── tests
    ├── general                             # General tests applied to all images
    ├── jupyterlab-cpu                      # Test applied to a specific image
    └── jupyterlab-tensorflow

Troubleshooting

If running using a VM and RStudio image was built successfully but is not opening correctly on localhost (error 5000 page), change your CPU allocation in your Linux VM settings to >= 3. You can also use your VM's system monitor to examine if all CPUs are 100% being used as your container is running. If so, increase CPU allocation. This was tested on Linux Ubuntu 20.04 virtual machine.

pablodecm/kubeflow-containers