MLOps

MLOps in prokube covers the path from interactive model development to a model served in production.

Use Labs for interactive work, then move repeatable or shared workloads into the MLOps services below.

Main Workflows

Workflow	Use
Pipelines	Run reproducible, inspectable workflows with Kubeflow Pipelines. Use this when notebook work should become a cluster-executed workflow.
Hparam Search	Run Katib experiments to compare parameter combinations and parallelize trials on cluster compute.
MLflow	Track experiments, log artifacts, manage model registry entries, and issue credentials for notebooks, pipelines, and serving.
Model Serving	Deploy trained models as KServe inference endpoints, including models from S3-compatible file storage or MLflow.
Model Serving Autoscaling	Tune KServe autoscaling with KPA, HPA, and KEDA, including vLLM token-throughput metrics.
Serverless	Run generic HTTP containers with Knative Serving when the workload is not a model-specific KServe deployment.

Typical Flow

1. Develop code and inspect data in a JupyterLab, VS Code Lab, or RStudio Lab.
2. Track experiments and model artifacts in MLflow.
3. Convert repeatable training or evaluation steps into Pipelines.
4. Use Hparam Search when parameters should be explored systematically across cluster resources.
5. Deploy the selected model through Model Serving.

Foundation

MLOps workloads run on the shared prokube Foundation. Start with these cross-cutting pages when you need platform behavior rather than tool-specific usage:

• Workspaces
• Kubernetes Resources
• Observability
• API Keys