CNCF Atlas

English

日本語

English

日本語

Appearance

Getting Started ​

Verified against KFP 2.16.1. Commands assume a running Kubernetes cluster, kubectl, and Python with pip.

Prerequisites ​

  • A Kubernetes cluster and kubectl configured against it.
  • Python 3 with pip for the kfp SDK.
  • This is the standalone deployment with no authentication, intended for trying KFP, not production.

Install ​

Deploy KFP standalone with Kustomize:

bash
export PIPELINE_VERSION=2.16.1
kubectl apply -k "github.com/kubeflow/pipelines/manifests/kustomize/cluster-scoped-resources?ref=$PIPELINE_VERSION"
kubectl wait --for condition=established --timeout=60s crd/applications.app.k8s.io
kubectl apply -k "github.com/kubeflow/pipelines/manifests/kustomize/env/dev?ref=$PIPELINE_VERSION"

Install the SDK:

bash
pip install kfp==2.16.1

A first working setup ​

  1. Port-forward the UI and API so the SDK can reach the API server.

    bash
    kubectl -n kubeflow port-forward svc/ml-pipeline-ui 8080:80

The UI is then reachable at http://localhost:8080.

  1. Author, compile, and submit a minimal pipeline.

    python
    from kfp import dsl, compiler, Client


@dsl.component
def say(msg: str):
    print(msg)


@dsl.pipeline(name="hello")
def hello_pipeline(text: str = "hi"):
    say(msg=text)


compiler.Compiler().compile(hello_pipeline, "hello.yaml")
Client(host="http://localhost:8080").create_run_from_pipeline_package("hello.yaml")

Verify it works ​

Open http://localhost:8080 and confirm the run appears and reaches a succeeded state, or check that the KFP pods are ready:

bash
kubectl -n kubeflow get pods

Where to go next ​

For authenticated, multi-user, and production deployments, follow the official KFP standalone installation guide and the Kubeflow Pipelines overview. For local backend development, the repository's CLAUDE.md documents a Kind-based deployment with make -C backend kind-cluster-agnostic.