Machine learning operations
Definition
Machine learning operations
Machine learning operations is the discipline of keeping a trained model working once it is live: versioning, pipelines, monitoring for drift, retraining triggers and rollback. It is the running of a model, not the building of one or the buying of one.
Practitioners shorten it to MLOps. The name borrows from software operations, and the borrowing is deliberate — a model in production is a running service with a decay problem attached to it.
OA’s entry on machine learning outsourcing answers the sourcing question, who builds and trains the model and under what contract. This page answers what has to exist afterwards so the thing keeps earning its place.
A provider who hands over trained weights and no pipeline has delivered an artefact — not a service.
The gap shows up fast — the first retrain nobody owns, the first quiet accuracy drop, the first rollback with no stored version to roll back to.
Key takeaways
- MLOps covers a model’s working life, not its construction or its purchase.
- Versioning, monitoring and a retraining trigger are the minimum viable set.
- Models decay quietly, so an automated drift alert beats a quarterly review.
- Ask a provider for the pipeline and the registry, not only for the weights.
How it works
MLOps runs as a loop rather than a launch. Data goes in, a pipeline trains and validates a candidate, a registry records the version, monitoring watches live predictions, and a breached threshold sends the loop round again.
Google Cloud’s MLOps guidance calls the starting point level 0, where every step is manual and driven by notebooks and a new model version reaches production only a couple of times a year.
Level 0 also splits the data scientists who build from the engineers who serve. The same guidance warns that this handover produces training-serving skew and leaves prediction behaviour unlogged, so performance degradation goes undetected.
| Component | What it gives you | What breaks without it |
|---|---|---|
| Model and data versioning | A reproducible pair you can name | Nobody can rebuild last month’s result |
| Training pipeline | A run anyone on the team can start | Retraining lives in one person’s notebook |
| Model registry | One place production pulls from | No clear answer to what is serving now |
| Live monitoring | Drift and accuracy alerts | A business metric finds the decay first |
| Retraining trigger | A rule rather than a reminder | Updates happen when somebody remembers |
| Rollback path | A tested previous version | An incident turns into a rebuild |
Governance language has caught up with the practice. The published AI Risk Management Framework, AI RMF 1.0 (2023), asks that post-deployment monitoring plans be implemented with mechanisms for user input, appeal and override, incident response and decommissioning.
Cost sits in the loop too — retraining is compute you pay for on a schedule you chose, so the trigger threshold is a budget decision as much as a quality one.
Ownership varies with size, so smaller teams give the loop to a data science lead who still writes models, while larger ones appoint an AI operations manager accountable for uptime, cost and the retraining calendar.
Examples
MLOps looks different by sector, but the tell is always the same: who gets paged when accuracy falls. Three shapes recur across engagements, and each one draws the line between building and running somewhere different.
Fraud scoring at a digital bank. The model retrains weekly because attackers change behaviour faster than anything else in the business, and the trigger is a drop in precision at fixed recall, not a date in the calendar.
Churn prediction at a telco. Marketing acts on the scores, so the operations team publishes a monthly stability report and freezes the model during a tariff change, when training data stops describing the world customers are living in.
Document extraction in shared services. Offshore finance teams run extraction models over supplier invoices. When one client changes its invoice template, accuracy falls for that client alone, so monitoring is sliced by customer rather than reported as a single average.
The common failure is the pilot that never grows up. Teams treat moving an AI pilot to production as a deployment task when it is mostly an operations build, and buyers of data science outsourcing inherit that gap at handover.
Related terms
These neighbouring terms sit closest to machine learning operations. Each answers a question this page deliberately does not, from who supplies the model in the first place to how the organisation decides what is worth automating.
- AI governance framework: the rules and roles above the work; MLOps is the running of one model inside them.
- AI use case discovery: choosing what to build; MLOps begins once that choice has shipped.
- Machine learning outsourcing: buying the capability to build a model; here the question is who keeps it alive.
- AI operations manager: the named person accountable for the loop this page describes.
- AI pilot to production: the one-off crossing; MLOps is the repeatable version of it.
FAQ
Is machine learning operations the same as DevOps?
No. The code side is familiar, but models also depend on data that shifts after release. MLOps adds versioned datasets, drift monitoring and retraining to the usual build and deploy chain.
How often should a model be retrained?
Retrain on a trigger rather than a timetable. Agree a metric, a threshold and a held-out test set in advance, then let a breach start the run.
Can machine learning operations be outsourced?
Yes, and it often is, usually priced per model per month as a run service. Insist that pipelines, registry access and monitoring dashboards belong to you rather than to the provider’s internal tooling.
What is model drift?
Drift is the gap that opens between the world a model was trained on and the world it now scores. It shows up as falling accuracy, shifting input distributions, or both at once.
What should a buyer ask for in an MLOps contract?
Ask for the pipeline, the retraining trigger and the rollback plan, with a named owner against each one. Agree the accuracy metric and the test set before work starts, because renegotiating them after delivery never goes well.
Compare providers that run models rather than hand them over in the Outsource Accelerator provider directory.
Independent
