codrinbucur / os-mlops

OS MLOps

This repository contains a number of assets for implementing an open source MLOps approach using OpenShift Container Platform and Red Hat OpenShift Data Science (RHODS).

RHODS Pipelines and Serving

• RHODS Workbench Controller based on Kubeflow Notebook Controller for notebook-based development,
• Data Science Pipelines based on Kubeflow Pipelines for workflow orchestration and experiment tracking,
• RHODS Model Serving based on KServe ModelMesh for model serving.

RHODS Demo Pack

To quickly set up a RHODS environment for fraud detection and object detection demos, follow these steps:

1. (optional) Set up GPU enablement if GPUs are present in your cluster.
2. Deploy the RHODS operator on your OpenShift cluster. The demo pack has been tested with RHODS 1.28.1.
3. Deploy the OpenShift Pipelines operator. The demo pack has been tested with Pipelines versions 1.8 to 1.10.
4. Deploy the Codeflare operator. The demo pack has been tested with Codeflare versions 0.0.4 to 0.0.6.
5. Clone this repository and navigate to manifests.
6. Run oc apply -f projects.yaml
7. Run oc apply -k .

Once the manifests have been deployed, your environment contains:

• A Minio instance as a lightweight S3 storage provider. You can manage the S3 buckets through the Minio UI through the minio-ui route URL in project minio. Use minio and minio123 for logging in.
• A Data Science Project fraud-detection for running the fraud detection demo. The pipeline server is instantiated and cluster storage and data connections are configured.
• A Data Science Project object-detection for running the object detection demo. The pipeline server is instantiated and cluster storage and data connections are configured. The OVMS model server is instantiatend for model deployment.
• A Data Science Project huggingface-demo for running the Huggingface demo. Cluster storage is configured.
• A number of community workbench images.
• A number of custom serving runtimes.

To get started with your demo, instantiate the respective workbenches.

Fraud detection

1. In the RHODS dashboard, enter the fraud-detection project.
2. Create a new workbench with an arbitrary name and these parameters:
   • image: Trino
   • existing cluster storage: development
   • existing data connection: fraud-detection
3. In the workbench, clone this repository, navigate to notebooks/fraud-detection and follow the instructions.

Object detection

1. In the RHODS dashboard, enter the object-detection project.
2. Create a new workbench with an arbitrary name and these parameters:
   • image: Object detection
   • existing cluster storage: development
   • existing data connection: object-detection
3. In the workbench, clone this repository, navigate to notebooks/fraud-detection and follow the instructions.

Huggingface demo

1. In the RHODS dashboard, enter the huggingface-demo project.
2. Create a new workbench with an arbitrary name and these parameters:
   • image: CodeFlare Notebook
   • existing cluster storage: development
3. In the workbench, clone this repository, navigate to notebooks/codeflare-examples/guided-demos and follow the notebooks.

References

• AI on OpenShift
• RHODS Documentation