Glossary
What is MLOps?
In short
MLOps (machine learning operations) is a set of practices and tools that move a machine learning model from an experiment to stable daily work: building, releasing, monitoring and retraining.
Machine learning · machine learning operations
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine you grew a tomato on the windowsill. Growing it once is not enough: you have to water it, make sure it doesn’t wilt, and move it to a bigger pot when it gets crowded. Smart programs are the same. You can make a program quickly, but you have to look after it every day. The rules for that care are what this work is about.
In plain words
Training a model on a laptop is only half of the job. Then you have to package it, put it on a server, connect it to work systems, watch for errors and retrain it on fresh data from time to time. If you do this by hand, every step depends on one person, and a month later nobody remembers which data built the current version. MLOps turns these steps into a pipeline: data, training, checks and release follow the same rules, and every version is recorded. Ordinary software development releases updates automatically in the same way.
For business
Without this discipline, models don’t live long: the author leaves, the data changes, accuracy drops, and someone notices a quarter later. A paper by Sculley and co-authors at NeurIPS 2015 called these hidden costs the technical debt of machine learning systems and warned of large ongoing maintenance costs. The budget covers storage of data and model versions, automatic training and checks, monitoring of quality and data drift, and on-call duty. For one small model, simple scripts and a written procedure are enough; ten models need a platform and a separate role. Before launch, ask the contractor who owns the code, the data and the model, how to roll back a failed version, and who watches quality after handover.
For a specialist
MLOps (machine learning operations) applies DevOps practices to machine learning systems, where data and models are versioned along with code. Kreuzberger, Kuhl and Hirschl (2022) reviewed the literature and tools and interviewed experts to describe its principles, components, roles and architecture. A typical setup has a data pipeline with schema checks, a feature store, experiment tracking, a model registry, an automated training pipeline, canary or shadow deployment, and monitoring of quality and drift. The Google Cloud guide (updated 28 August 2024) names three maturity levels: 0, a manual process; 1, an automated training pipeline; 2, automated CI/CD of the pipeline itself. It adds continuous training (CT) to CI and CD: automatic retraining and release of models. The same guide notes that model code is only a small part of a real system. Sculley and co-authors (NeurIPS 2015) listed the sources of technical debt: eroded boundaries, entanglement, hidden feedback loops, undeclared consumers, data dependencies and configuration issues.
Example
A retailer runs twelve demand forecasting models, one per region. An analyst used to retrain each of them by hand once a quarter. Now the pipeline does it every week. A new version goes live only if it beats the old one on fresh data, and if quality drops, the system rolls it back and alerts the person on duty.
Related terms
Questions and answers
Sources
- Sculley et al., Hidden Technical Debt in Machine Learning Systems, NIPS 2015papers.nips.cc
- Kreuzberger, Kühl, Hirschl, Machine Learning Operations (MLOps): Overview, Definition, and Architecture, 2022arxiv.org
- Google Cloud, MLOps: Continuous delivery and automation pipelines in machine learning, updated 28 August 2024docs.cloud.google.com
Where companies use it
Not sure whether your company needs this?
AI solutions map