Kubeflow Pipelines SDK (KFP SDK)

To create a KFP SDK pipeline:

Ensure your Provider supports KFP SDK by specifying the KFP SDK image in spec.frameworks[].
Create a Pipeline resource, specifying:
- the kfpsdk framework in spec.framework.name. This needs to match the name specified in the Provider.
- the fully qualified name of the Python function that creates a KFP SDK pipeline under spec.framework.parameters[].pipeline.

KFP SDK Parameters

Name	Description
`pipeline`	Fully qualified name of the Python function creating a KFP SDK pipeline. This function should be wrapped using the `kfp.dsl.Pipeline` decorator.

Environment Variables

The KFP SDK compiler automatically injects the following environment variables during compilation:

Variable	Description
`KFP_PIPELINE_IMAGE`	Set to the value of `spec.image` from the Pipeline resource. Use this in your pipeline code to dynamically set component base images, e.g. `@dsl.component(base_image=os.environ.get("KFP_PIPELINE_IMAGE"))`.

KFP SDK Pipeline resource example

apiVersion: pipelines.kubeflow.org/v1beta1
kind: Pipeline
metadata:
  name: kfpsdk-quickstart
spec:
  provider: provider-namespace/kfp
  image: kfp-operator-kfpsdk-quickstart:v1
  framework:
    name: kfpsdk
    parameters:
      pipeline: getting_started.pipeline.add_pipeline

Component naming

The COMPONENT used in a Run’s artifact path is derived from the @dsl.component Python function name, which the KFP SDK normalizes: the name is lower-cased and every run of non-alphanumeric characters (e.g. underscores) is replaced with a single hyphen. CamelCase is not split into words (e.g. a function named BatchPredictor becomes batchpredictor, not batch-predictor).

Pipeline code (function name)	COMPONENT to use
`def pusher(...)`	`pusher`
`def train_model(...)`	`train-model`
`def BatchPredictor(...)`	`batchpredictor`