MLOps: Continuous Delivery and Automation Pipelines in Machine Learning

What this guide is

This is an article in the Google Cloud Architecture Center that sets out how to apply DevOps practices to machine learning systems. It has become one of the most referenced descriptions of MLOps, largely because of its three maturity levels, which many teams and courses use as a shared vocabulary.

The three levels in brief

• Level 0, manual process: every step is manual, from data preparation to training and validation. Data scientists hand a model to engineers, releases are rare and there is no CI/CD.
• Level 1, ML pipeline automation: the training pipeline is automated, so the model is retrained in production on fresh data when triggered. This is continuous training, and it brings data validation, model validation and metadata management with it.
• Level 2, CI/CD pipeline automation: the pipeline itself is built, tested and deployed automatically, so new ideas move from experiment to production quickly and reliably.

Why it is worth an hour

It explains why ML needs more than normal CI/CD. Code is only one of the things that change: data changes, models decay, and a system must test data and models as well as code. The guide names continuous training as a property unique to ML systems, and lists the components a mature setup needs, such as a feature store, a metadata store and a model registry. Reading it helps you place any tool you learn into a bigger picture, and helps you explain your own project's maturity honestly.

How to read it

Read it once early, to see where the lifecycle is heading, and again when you design CI/CD for your own project. Ask at each level what your project automates today and what single change would move it up. Ignore the product-specific parts if you do not use Google Cloud: the ideas map onto open source tools such as MLflow, DVC, Prefect or Airflow and GitHub Actions.

Prerequisites

Basic knowledge of how a model is trained and evaluated, and what CI/CD means for normal software.

Free or paid?

Free to read. The page shows it was last updated in August 2024; the concepts have aged well, while specific Google Cloud product names may change.

Common questions

Does my project need level 2? Rarely at first. Most teams get the biggest gains from moving level 0 to level 1 with a gated, automated training pipeline.

Does it cover LLM applications? Not directly. The same thinking applies, with prompts, eval sets and indexes as the moving parts.

What can I do after it? Describe your project's MLOps maturity, and plan the next automation step with a clear reason.

Best for: ML learners and engineers who want the big picture of MLOps before, or while, learning the tools.