Skip to content

glossary

MLOps

The practice of applying DevOps principles to machine learning: versioning data and models, automating training and deployment pipelines, and monitoring models in production for drift and degradation.

In depth

MLOps extends CI/CD thinking to systems whose behavior comes from data as much as code. A machine learning system has more moving parts than ordinary software: training data, feature pipelines, model code, hyperparameters, and the trained model artifact itself, and all of them must be versioned and reproducible. MLOps pipelines automate the path from raw data to deployed model: data validation, feature engineering, training, evaluation against quality gates, registration in a model registry, and deployment as an API or batch job. Crucially, the work does not end at deployment. Production models degrade silently as the real world drifts away from the training data, so MLOps includes monitoring prediction distributions, data quality, and business metrics, with automated retraining or alerts when drift is detected. Experiment tracking tools record every run so results are comparable and auditable. The goal is to make shipping and maintaining model number fifty as routine as shipping model number one.

Why it matters

Most ML projects historically failed not at modeling but at productionization, with models stuck in notebooks or silently rotting in production. MLOps turns ML from artisanal experiments into a dependable engineering discipline, which is why MLOps engineers are among the most sought-after hybrid roles.

Real-world example

example.txt

A bank's fraud-detection team uses MLflow to track experiments and a pipeline that retrains weekly on fresh transactions. Evidently monitors detect that a new payment method has shifted the input distribution, prediction confidence drops, and the pipeline automatically triggers retraining and a canary deployment of the new model, which restores precision before fraud losses grow.

Tools related to MLOps

MLflowKubeflowVertex AIAmazon SageMakerDVCEvidently

Interview questions

  1. How does MLOps differ from traditional DevOps?
  2. What is data drift versus concept drift, and how do you detect each?
  3. How do you version a machine learning model and its training data?
  4. Design a CI/CD pipeline for a machine learning model.
  5. What is a model registry and why do teams use one?
  6. How would you safely roll out a new model version to production?