Pipeline Frameworks

Overview

The KFP Operator supports multiple pipeline frameworks, including:

• TFX

Compiler Workflow

The kfp-operator-create-compiled workflow compile step accepts the following parameters:

• resource-image: the image containing the model code (looked up from the pipeline resource in preview workflow step)
• provider-config: the provider configuration (see provider overview)
• pipeline-framework-image: the image for the pipeline framework compiler

The image specified in pipeline-framework-image is executed as a initContainer and runs the entrypoint script. The entrypoint script should copy the required compiler code into the shared directory /shared (which is mirrored into the main container) and then exit. This /shared location is passed as the first and only parameter to the entrypoint script.

Once the init container has complete then the main container is executed. The /shared/compile.sh which needs to be provided by the compiler image should simply execute the compiler module.

See examples of entrypoint and compile scripts here.

Adding a Custom Pipeline Framework

Firstly a Docker image needs to be created that contains the necessary dependencies for the custom pipeline framework. This image should be pushed to a container registry that the KFP Operator deployment has access to. e.g. ghcr.io/kfp-operator/kfp-operator-custom-compiler:version-tag

The Docker image needs to conform to the correct structure for the KFP Operator to be able to use it. See readme for more information.

As seen in configurations, available frameworks to the KFP Operator installation are defined in the frameworks field of the operator configuration. So for example:

    framworks:
      - name: tfx
        image: ghcr.io/kfp-operator/kfp-operator-tfx-compiler:version-tag
      - name: custom
        image: ghcr.io/kfp-operator/kfp-operator-custom-compiler:version-tag

Then to use the custom framework in a pipeline simply configure the framework attribute in the pipeline resource.