DVC + Docker example

A few different ways to run DVC with Docker.

• inside - the whole DVC pipeline runs inside a container.
• outside - that we run docker run as one or more stages of a DVC pipeline while dvc repro or dvc exp run running outside a container.

Don't consider this as comprehensive tutorial. It's provided as an example of different combinations. There are other ways to manage different aspects of this scenario, reach out to us or read the docs to get the right combination of commands or the right way to package things into an image.

inside/mount

Mounts a current directory (workspace) as volume to an running image. The workflow to run it. Suits well if you need to run it locally since it takes the changes in the workspace, local Git config. Also, it avoid clones, duplicating data and results to pass them between the container and a host.

1. do any changes to the wokspace, code, data, etc
2. run any DVC commands if needed - dvc pull, etc to get data
3. run DVC pipeline with:

docker run -it \
           -v $(pwd)/../..:/app/workdir \
           -v ~/.gitconfig:/etc/gitconfig \
           dvc-docker-mount

On Mac metal add --platform linux/amd64 since DVC deb package is not available at time of writing this.

inside/clone

Suits well for the scenario when we need to run an experiment, or make an ELT remotely - on demand or on schedule. We do a fresh git clone, pull data, run the pipelein (can be dvc repro ro dvc exp run) and push results back- either to be consumed within Studio (it renders dvc exp push-ed experiments), or by git clone, dvc pull, dvc exp pull commands on any other machine.

To run it:

docker run -e STUDIO_GIT_TOKEN='*********' \
           -e AWS_ACCESS_KEY_ID='*******' \
           -e AWS_SECRET_ACCESS_KEY='***********' \
           -it dvc-docker-clone

Where:

• STUDIO_GIT_TOKEN - a secure way to give access to a specific repo for a certain period of time. Read more here. Other platforms might not provide this. In this case you can use SSH key and/or create a seprate account with a restricted access to make it secure.
• AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY - an example on how to pass AWS credentials to pull / push data if needed. Again, it depends on the environment if you need this or not. On an EC2 instance we would recommend to setup an IAM role and avoid managing credentials manually.

On Mac metal add --platform linux/amd64 since DVC deb package is not available at time of writing this.

outside

Run with dvc repro or dvc exp run. It shows a way to package a Dockerized commands into a DVC pipeline. An additional benefit is that DVC handles the Docker image update as well.

Considerations

1. Storage and Git credentials. Since we need to run the pipeline or certain commands we need to pass Git credentials and/or storage credentials + Git config. See some comments in the ``inside/clone` section that gives a bit more color to this.
2. Install DVC. We use the deb package and Ubuntu base image in these examples. A clear optimization is to create your own image that already includes all the basic steps and can be reused in the projects' repos to simplify their Dockerfiles.
3. Iterative Studio - gives an excellent way to see the results, trigger runs, register and see all the models, and more.