Architecture Asset: MLOps Training and Serving on STACKIT

Overview

This pattern operationalizes machine learning. Experiments run in notebooks, training executes on Kubernetes, models are versioned in an Object Storage registry, and the selected model is served behind a load balancer with monitoring and retraining loops.

Typical use case

• Reproducible ML: version data, code, and models across the lifecycle.
• Automated retraining: retrain and revalidate as data changes.
• Reliable serving: deploy models with rollback and monitoring.

Architecture diagram

Design best practices

• Automate the lifecycle: use CI/CD for models, not manual promotion.
• Version everything: data, features, code, and models for reproducibility.
• Monitor drift: feed monitoring signals back into retraining.
• Roll out safely: support canary or shadow deployment and fast rollback.