Skip to content

Chapter 1.10 — MLOps, DataOps and FinOps applied to ML

🎯 Objective

Position the three operational disciplines that sustain ML in production and show why all three must coexist. Handling one without the others is the classic pattern of a team that delivers a good model in the lab and then gets lost when it enters continuous operation.

🧠 Concepts

  • MLOps — CI/CD for models, pipelines, observability, governance. Treats the model as a versioned artifact with its own lifecycle.
  • DataOps — quality, lineage, contracts, freshness, SLAs. Treats data as a product, with an owner and an SLA.
  • FinOps — cost visibility, allocation and optimization. Treats cost as an operational metric, not as a surprise at the end of the month.

In practice, these three disciplines overlap at critical points. A drift detected by MLOps may trigger retraining that needs DataOps approval (new dataset quality) and generates extra consumption that FinOps must absorb. Mature teams treat the three as a single operational loop, not as parallel squads.

🏗️ Patterns

  • Reproducible pipelines (Airflow, Dagster, Prefect, Kubeflow).
  • Feature store with offline/online parity.
  • Model registry with approval.
  • Continuous monitoring with SLO/SLI alerts.
  • Model catalog with owner, risk, status.

📌 Checklist

  • [ ] Are training, validation, deployment and rollback reproducible?
  • [ ] Is there a model catalog with owners?
  • [ ] Is there a budget per model/tenant?