Introduction

This demo shows how KFP SDK can be used to create the Data Science Pipeline in OpenShift AI.

Important Disclaimer

IMPORTANT DISCLAIMER: Read before proceed!

1. These examples are to showcase the capabilities of OpenShift AI.
2. Certain components may not be supported if they are using upstream projects.
3. Always check with Red Hat or your account team for supportability.

Requirements

• Tested on OpenShift 4.16/4.17 with cluster-admin

• Tested on OpenShift AI 2.17 with Model Registry set to Managed inn DataScienceCluster default-dsc:

  spec:
    components:
      modelregistry:
        managementState: Managed
        registriesNamespace: rhoai-model-registries

• OpenShift Service Mesh is installed

• OpenShift Severless is installed

• Authorino is installed

Setup

This demo will be installed in the dsp-demo-creditcard-fraud namespace.

make setup-dsp-demo

Additionally, it will install the following components:

Component	Namespace	Description
Gitea	gitea	Store our pipeline code
OpenShift Pipeline	dsp-demo-creditcard-fraud	To build our Data Science Pipeline using KFP SDK
Model Registry	rhoai-model-registries	To store the model meta information
MariaDB	rhoai-model-registries	MySQL backend for model registry
Data Science Pipeline	dsp-demo-creditcard-fraud	To run the pipeline
Minio	dsp-demo-creditcard-fraud	To store the model and pipeline artifacts

To teardown the entire setup. This will delete the namespace and the model registry.

make teardown-all

Minio

The credential for minio is minio/minio123. Route for minio is:

echo "https://$(oc get routes minio-console -n dsp-demo-creditcard-fraud -o jsonpath='{.spec.host}')"

Gitea

The credential for opentlc-mgr admin user in gitea can be found in

oc get gitea/gitea -n gitea -o jsonpath='{.status.adminPassword}'

Open Data Hub Tools & Extensions Companion

This includes a S3 browser.

make setup-odh-tec 

Running the Demo

You can trigger the pipelione build using 2 methods:

• Gitea has a webhook to trigger the Tekton pipeline.
• Run the pipeline run script

make run-pipeline

1. Tekton will be triggered to build the Data Science Pipeline

   

2. The Data Science Pipeline will run to train the model

   

   

3. Model will be saved to S3 bucket

   

4. Model will be registered with the model registry

   The model details:

   The various versions:

   The details of a model version. E.g. version 4:

5. Model will be deployed to KServe

   

Deploying the model using model registry

After the model is registered, you can manually deploy it.

1. Manual deployment

2. Model registry deployment view

3. The InferenceService annotations

Known Issues

RHOAI Admin

To add kube:admin to RHOAI admin group

Always the Latest Model Version

When you use the model registry to deploy, the current latest model will always be deployed. This is due to how OVMS and KServe storage-initializer works.

OVMS requires a specific model repository structure. KServe uses the model path as the S3 path prefix to download all the files into the container.

model_config.json was an attempt to solve this using a custom ServingRuntime as part of the data science pipeline. However, when you deploy using the model registry, the ServingRuntime is the default template version.

Previous Model Accuracy Metric

Is currently hardcoded for now.

Credits

Folks who helped me start this repo and provided guidance:

• Trevor Royer @strangiato
• Humair Khan @HumairAK
• Matteo Mortari @tarilabs
• Yuan Tang @terrytangyuan

To-Do

1