Skip to main content

Defining Custom Images

All tasks executed remotely run in a container, both on Kubernetes and in AWS Batch. Hence, the default environment for remote tasks is defined by the container (Docker) image used.

By default, Metaflow uses a default Python image which doesn't contain any libraries besides Python itself. When additional libraries are needed, an easy option is to use the @pypi and @conda decorators which install libraries on the fly, on top of the base image.

Alternatively, you can use any other image of your choosing. Many off-the-shelf images work with Metaflow without modifications, or you can build a custom image.

Building a custom image

The main requirement is to make sure that the image has a python version installed.

For more information about building a custom image, see this external howto page.

Configuring a custom image

There are three ways to pick an image, depending how broadly you want to use the image. The options are listed from the broadest to the most specific:

1. Define a default image

If you want to use an alternative image by default in all remote tasks, specify two variables in the Metaflow configuration files:

  • METAFLOW_DEFAULT_CONTAINER_REGISTRY controls which registry Metaflow uses to pick the image. This defaults to DockerHub but could also be a URL to a public or private ECR repository on AWS.

  • METAFLOW_DEFAULT_CONTAINER_IMAGE dictates the default container image that Metaflow should use.

2. Define a step-specific image

To use a specific image in a specific step, specify the image argument in the @batch or @kubernetes decorators.

3. Execute a run with a custom image

You can test a specific image with a run without changing anything in the configuration or the code. Simply add :image=repo/image:version to --with batch or --with kubernetes.

For instance, you can run with the latest Ubuntu image like this:

python run --with

Custom image with @pypi or @conda

You can use both a custom image as well as @pypi or @conda on top of it. As usual, @pypi and @conda guarantee isolated environments, meaning that packages installed in the image won't be visible in steps unless you explicity disable the environment for a step.

This combination is beneficial if there are other assets in the image besides packages that should be accessed by steps. Steps may access all files in the image, e.g. configuration files, background processes work as usual, and you can launch image-specific subprocesses.

This way, you can design a base image containing non-library assets, and let developers handle libraries independently with @pypi or @conda.