MLOps with Openshift AI

About

This demo illustrates MLOps features of Openshift AI and extensions to Openshift (streaming, ELT, ...).

Content

• Openshift AI MLOps features (notebooks, automated pipeline creation through git, experimentations, artifacts, model registry)
• Streaming with kafka
• Data transformation with camel
• Kserve features (usage of transformer for pre/post processing)

Deployment

The following procedure will deploy all the demo components. If you want to deploy only specific components, look at this documentation.

Operators

Install the operators.

oc apply -k ./manifests/operators/

Wait for the installations to complete. Confirm that all operators are ready.

Operator instances

Deploy the data science cluster and the knative instances by runnning:

oc apply -k ./manifests/operators-instances/

Demo environment.

Deploy the demo instances:

helm template ./manifests/instances/core | oc apply -f -
oc kustomize ./manifests/instances/automated-pipelines/ --enable-helm | oc apply -f -
oc kustomize ./manifests/instances/streaming/ --enable-helm | oc apply -f -

Required manual setup on Openshift AI

1. Data science pipeline

Go to the digit-recognition data science project. On the pipeline tab, create a new pipeline server using one of the data connection available. Change the bucket name to "ml-pipelines" for clarity. Wait for the pipeline server creation to complete.

2. Notebook creation

Create a new workbench. Use the standard data science container image. Add the "data" data connection. Wait for the notebook creation. Open it and clone https://github.com/adrien-legros/rhods-mnist-model.

3. Model registry setup

On the settings side of Openshift AI choose model registry. Create a new one with the settings:

• Name: model-regisgtry
• Host: mysql.mnist
• Port: 3306
• Database name: modelregistry

Wait for the model registry creation to complete.

Credentials: username / password

• Openshift AI: your openshift user (needs to be a openshift ai admin)
• Gitea: data-scientist-1 / rhods
• Grafana: admin / admin

[UNCOMPLETE] Architecture

The schema does to reflect:

• Model registry
• Experimentations and metric tracking
• Kserve as a variation for the streaming workflow

[DEPRECATED] Walkthrough

Deprecated but gives some steps for the demo. Walkthourgh and highlights can be found on this documentation.