manifold/orbyter-docker

Manifold's Orbyter Docker image toolset helps data science teams easily move to a container-first workflow, from local development to serving in production settings. The goal of these tools is to bring DevOps and software engineering best practices to the data science community to increase productivity and quality of delivered work to customers (internal and external). Several Dockerhub repositories are represented in this code repository: manifoldai/docker-ml-dev, manifoldai/docker-dl-dev, ...., all of which are listed below.

General structure

Each docker image (dockerhub repo) has a folder with the following file structure:

orbyter-docker/
├── docker-repo/          # e.g, docker-ml-dev
│   ├── Dockerfile        # Commands for building this image
│   ├── OVERVIEW.md       # Dockerhub image description
│   ├── README.md         # Additional image information
│   ├── VERSION           # See below
│   └── requirements.txt  # Packages unique to this image
└── requirements-dev.txt  # Packages installed on all dev images

All development images (i.e. non-base images) first install packages from requirements-dev.txt using pip. Packages from the directory-specific requirements.txt are then installed.

Versioning and tagging

Base images use major/minor (A.B) versioning.

Development images have A.B.C versioning to indicate that base image version A.B was used. The requirements.txt file

The tip of master contains the latest version of each Docker image (VERSION). Each repo VERSION is given a specific tag. To audit/compare previous versions, you need to reference the image by its git tag.

Each given image release is marked by a git tag. The tagging convention is <image-name>-<version>. For example, the git tag tag:orbyter-ml-dev:5.1.2 corresponds to version 5.1.2 of the image orbyter-ml-dev and it was built upon orbyter-base-sys:5.1. Pushing a tag to GitHub will trigger a deployment of that image to Dockerhub.

Image release steps

We'll describe how to release a new orbyter-ml-dev, but the steps are the same for other images. Let's say we are releasing version 5.1.2 of orbyter-ml-dev.

1. Create a new branch, e.g, gg/patch-mlflow-5.1.2
2. Make changes to Dockerfile, requirements.txt, README.md, etc. and bump VERSION to 5.1.2. Note: you can test your new build by running make build target=orbyter-ml-dev from the top-level directory
3. Create a pull request back into master.
4. When changes are merged to master, get the latest commit: git checkout master, git pull
5. Tag the code: make release target=orbyter-ml-dev. This will push the tag to origin and kick-off a GitHub actions job that will push two new images to Dockerhub: manifoldai/orbyter-ml-dev:5.1.2 and manifoldai/orbyter-ml-dev:latest.

Patching an image

In certain cases it may be necessary to delete a tag and add it again, retriggering deployment to Dockerhub.

# delete tag locally
git tag -d orbyter-image-name-1.1.1
# push deletion to github
git push origin :refs/tags/orbyter-image-name-1.1.1
# retag to trigger deployment
make release target=orbyter-image-name

Docker images

orbyter-base-sys

Base docker image for machine learning development in python. ML images are built on top of this, and this should not be used directly for development. This base image contains basic tools like vim, emacs, curl, and python, but does not install any ML specific packages.

orbyter-base-sys-dl

Base docker image for deep learning development in python, which contains CUDA libraries. DL images are build on top of this, and this should not be used directly for development. This base image, like orbyter-base-sys, contains basic tools like vim, emacs, curl, and python. Because it contains the CUDA libraries, it is compatible with many deep learning frameworks like PyTorch, but the platform architecture is limited.

orbyter-ml-dev

Docker image for ML development in python which contains essential tools like jupyter, pandas, numpy, and scikit-learn.

orbyter-dl-dev

Docker image for CUDA/GPU-assisted deep learning development.

orbyter-spark-dev

Docker image for ML projects with Apache Spark.

Historical notes

Support for an MLFlow image was dropped. Use the official MLFlow image from Github Container Registry instead.

After commit e039c36, this repo was drastically reorganized. <=e039c36, each version of each docker repositories image was given as a subfolder. E.g,

orbyter-ml-dev/
├── 3.1
│   ├── Dockerfile
│   └── ...
└── 3.2
    ├── Dockerfile
    └── ...