EXPLAINER

MLOps (Machine Learning Operations)

conceptfreshness: evolvingsources not yet verified

MLOps applies software-engineering and operations practices — versioning, automation, testing, deployment, monitoring — to machine-learning systems so they stay reliable in production.

Short answer: MLOps = reproducible pipelines for data, training, evaluation, deployment and monitoring. It manages the extra moving parts of ML (data, models, experiments) beyond ordinary code. For LLM apps the same ideas appear as "LLMOps": prompt/version management, evals and observability.

What is it?

Building a model once is the easy bit. Keeping it working — retraining when data changes, tracking which version is live, noticing when quality drops — is the job MLOps covers.

Also known as: MLOps, LLMOps, machine learning operations, ML pipeline, model lifecycle, CI/CD for ML, model deployment pipeline, ML platform.

Technical definition

Sculley et al. (2015) describe hidden technical debt in ML systems: data dependencies, entanglement and feedback loops. MLOps practices include data and model versioning, experiment tracking, CI/CD for training and serving, feature management, automated evaluation gates, staged rollouts, monitoring for drift and performance (Model Drift and Production Monitoring), and governance artefacts (Model Cards and Datasheets (AI Documentation)). Maturity models (for example from cloud vendors) describe progression from manual to automated pipelines.

Entity type: concept · Freshness class: evolving.

When should I use it?

Relevant once a model affects real users or decisions.

Implementation patterns

  • Version data, code and models together.
  • Automate train → evaluate → register → deploy with approval gates.
  • Shadow or canary deployments.
  • Monitor inputs, outputs and business metrics.

Technologies and frameworks

Commonly associated: MLflow, Kubeflow, Ray, DVC, Airflow/Prefect, cloud ML platforms. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Containerise the environment.
  2. Track experiments.
  3. Add an evaluation gate.
  4. Register and version models.
  5. Deploy gradually and monitor.

Common failure modes

  • Notebook-only pipelines nobody can reproduce.
  • Training-serving skew.
  • No rollback path.
  • Monitoring only infrastructure, not model quality.

Troubleshooting

  • Cannot reproduce a result — Pin data/code/environment versions and seeds; use experiment tracking.
  • Production quality differs from offline — Check training-serving skew and drift.

Contrasted with

  • CI and CD — CI/CD handles code; MLOps also handles data, models and experiments.

Questions this page answers

  • What is MLOps?
  • What is LLMOps?
  • How do I deploy a machine learning model to production?
  • What is the difference between DevOps and MLOps?

Typical goals: move a model from notebook to production; organise ML team workflows.

Sources and verification

Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.

Freshness class: evolving — Established techniques whose best practice shifts as research and tooling move. Re-review every 6 months.

Canonical sources (unverified leads):

  • Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS (2015) — locator to be identified
  • Google Cloud, MLOps: Continuous delivery and automation pipelines in machine learning — cloud.google.com/architecture (locate current page)