ML Pipelines
PAIS provides a full ML pipeline stack for automating training, evaluation, and deployment workflows.
Components
| Component |
Purpose |
URL |
| Kubeflow Pipelines |
Define and run multi-step ML workflows as DAGs |
https://kubeflow.pais.auckland.ac.nz |
| Katib |
Automated hyperparameter tuning (Bayesian, random, grid search) |
Kubeflow → Experiments (AutoML) |
| MLflow |
Experiment tracking, model registry |
https://mlflow.pais.auckland.ac.nz |
| KServe |
Model serving (deploy a trained model as an inference endpoint) |
Kubeflow → Models |
When to use pipelines vs notebooks
| Scenario |
Use |
| Exploratory analysis, prototyping |
Notebook |
| Multi-hour training jobs |
Pipeline (runs unattended) |
| Hyperparameter sweeps |
Katib |
| Reproducible, versioned workflows |
Pipeline |
| Model serving post-training |
KServe |
Quick links