Skip to main content

AI glossary Operations

What is MLOps?

← All glossary terms
In short

The set of practices for deploying, monitoring and maintaining machine learning models in production in a reliable, repeatable way.

Explained

MLOps (machine learning operations) applies DevOps practices to machine learning: version control, automated pipelines, testing, deployment, monitoring and rollback, extended to data and models. The model code is only a small part of a production ML system; around it sit data collection and validation, feature pipelines, training, serving and monitoring. MLOps automates that whole chain so that models can be retrained, checked and released without manual steps. Without it, models that worked in a notebook tend to degrade or break once real data changes.

How it works

An MLOps pipeline turns model building into a repeatable process. Data is validated, the model is trained and evaluated, and if it passes the agreed thresholds it is packaged and deployed. In production, its inputs and outputs are monitored, and a drop in quality or a change in the data can trigger retraining or a rollback.

The main practices:

  • Versioning of code, data, features and models, so any result can be reproduced.
  • Continuous integration and delivery for both the code and the trained model.
  • Continuous training when new data arrives or performance drops.
  • Monitoring of quality, data drift, latency and cost.
  • A model registry with approvals, lineage and the option to roll back.

MLOps vs DevOps

DevOps deals with code that behaves the same way until someone changes it. A machine learning model can get worse without any code change, simply because the data it sees in production is different from the training data. MLOps adds what that requires: data validation, model evaluation, continuous training and monitoring of the model's behaviour, not only of the service.

MLOps for LLMs and generative AI

With large language models, the same ideas apply to new assets: prompts, system prompts, retrieval indexes, evaluation sets and guardrails. These need versioning, testing and monitoring just like a trained model. This is often called LLMOps. The emphasis moves from retraining towards evaluation, cost control and tracking answer quality over time.

Frequently asked questions

What is MLOps in simple terms?

It is the way of working that takes machine learning models from an experiment to a service that runs reliably every day. It automates training, testing, deployment and monitoring so that models stay useful as data changes.

What is the difference between MLOps and DevOps?

DevOps automates building and releasing software. MLOps adds what models need on top: versioned data, model evaluation, retraining and monitoring for drift, because a model can degrade without any change to its code.

Why do machine learning models fail in production?

Often because the data changes, the pipeline that feeds the model breaks or nobody notices the quality dropping. The model itself is a small part of the system; most failures come from the data and infrastructure around it, which is what MLOps addresses.

What is LLMOps?

LLMOps applies MLOps practices to large language models. Besides models, it versions and tests prompts, retrieval indexes and guardrails, and it monitors answer quality, latency and cost in production.

Which tools are used for MLOps?

There are open-source tools and managed services from cloud providers for pipelines, experiment tracking, model registries, serving and monitoring. The choice depends on the existing platform; the practices matter more than any single tool.

Decision guides that use this term

How to put it to work

Where it shows up at Thinkia

Products

Further reading

Sources

Updated: 2026-10-21

Want to see how this applies to your organisation?

Talk to an AI Expert