Installing Kubeflow Pipelines on OpenShift

Simplified instructions to install Kubeflow Pipelines (KFP) on OpenShift, running on OpenShift Pipelines (Tekton).

Prerequisites

OpenShift 4 Cluster
Admin access to cluster
OpenShift CLI

Prep

Install Open Data Hub

Login to the cluster using the console and admin credentials.
Navigate to OperatorHub and search for Open Data Hub Operator
Install using the beta channel. The version installed in this guide is 1.0.
Under the install operators, you should see the status of the operator change to Succeeded

Install OpenShift Pipelines (Tekton)

Login to the cluster using the console and admin credentials
Navigate to OperatorHub and search for OpenShift Pipelines
Pick the channel corresponding to your version of OpenShift
Install and verify the status of the operator changes to Succeeded

KFP on Tekton

Install

Follow these instructions to prepare your environment and download the tkn and kfctl binaries.

Deploy KFP

export KFDEF_DIR=<set-your-path>
mkdir -p ${KFDEF_DIR}
cd ${KFDEF_DIR}
export CONFIG_URI=https://raw.githubusercontent.com/IBM/KubeflowDojo/master/OpenShift/manifests/kfctl_tekton_openshift_minimal.v1.1.0.yaml
kfctl apply -V -f ${CONFIG_URI}

Open KFP console in your browser

oc expose svc ml-pipeline-ui -n kubeflow
echo $(oc get route ml-pipeline-ui -n kubeflow --template='http://{{.spec.host}}')

Smoke Test

Install the KFP-Tekton-SDK

Remember to add dsl-compile-tekton to your path. If you have both Python2 and Python3, you might have to add the binary to your path, e.g. export PATH=$PATH:/Users/.../Library/Python/3.8/bin

Make a local copy of the echo pipeline example as echo_pipeline.py.

Make a local copy of the sequential example as sequential.py.

Compile both examples

dsl-compile-tekton --py echo_pipeline.py --output echo_pipeline.yaml
dsl-compile-tekton --py sequential.py --output sequential.tar.gz

Open KFP console in your browser

echo $(oc get route ml-pipeline-ui -n kubeflow --template='http://{{.spec.host}}')

Upload pipelines echo_pipeline.yaml and sequential.tar.gz. For example,

For each pipeline, start a Pipeline Run in the Default Experiment.

If you navigate to your Default Experiment, you should see both pipeline runs complete.