MLOps

Why does MLOps matter?

Without MLOps, most model work stays stuck in notebooks or gets rebuilt by hand every time it reaches production.

In financial services, SR 26-2, the Federal Reserve's current model risk management guidance, requires banks to maintain a model inventory and apply validation and monitoring calibrated to each model's materiality. MLOps is the operational layer that makes that inventory and monitoring possible at the scale most banks now operate at, rather than something maintained separately from the models themselves.

In life sciences, models supporting drug discovery or clinical decision support need to be reproducible down to the software environment they ran in, since a regulatory submission may need to recreate a result years later. MLOps practices, like environment versioning and experiment tracking, are what make that kind of retroactive reproducibility realistic.

Stages of MLOps

What are the benefits of MLOps?

• Faster iteration: teams move through the manage-develop-deploy-monitor cycle without rebuilding infrastructure for every model.
• Governed scale: security, compliance, reproducibility, and auditability apply consistently across every model in production, including the ones built long after the first.
• Continuous performance: ongoing retraining and monitoring keep models accurate as the underlying data shifts.

How is MLOps different from DevOps?

DevOps and MLOps share the same emphasis on code, pipelines, and continuous collaboration between teams. The difference is that software behaves the same way in production as it did in testing, while a model's behavior can drift as the data it sees changes. MLOps takes DevOps principles and adds practices, like drift monitoring and retraining, to manage that changing behavior.

How is MLOps different from LLMOps?

LLMOps is MLOps applied specifically to large language models. The core lifecycle stages are the same, but LLMOps adds practices for prompt engineering, retrieval-augmented generation, and the much larger compute and evaluation requirements that come with foundation models.

Industry use cases

Financial services

Banks running dozens or hundreds of models in production need model monitoring and validation running on schedule across the full inventory, not tracked manually alongside the models themselves.

Life sciences

A drug discovery model published in a submission today may need to be rerun exactly, down to the package versions, if a regulator asks a follow-up question two or three years later, long after the original environment would otherwise have been upgraded or discarded.

Public sector

Government and defense deployments often run in hybrid or fully on-premises environments with strict data residency and security requirements. MLOps practices need to work the same way in an air-gapped or on-prem environment as they do in the cloud, since the operational discipline doesn't change just because the infrastructure does.

FAQ

What does MLOps stand for?

MLOps stands for machine learning operations, the operational discipline that keeps machine learning models reliable, secure, and reproducible from development through production.

Is MLOps the same as DevOps?

No. MLOps borrows DevOps practices like CI/CD and automated pipelines, but adds handling for the parts of the lifecycle unique to models, like data versioning, model drift, and retraining, that don't have a direct DevOps equivalent.

What tools are used for MLOps?

Common categories include experiment tracking (like MLflow), model registries, orchestration and pipeline tools (like Kubeflow or Airflow), and monitoring tools that watch for drift in production. Most enterprise MLOps setups combine several of these rather than relying on one tool for the whole lifecycle.

Do you need MLOps for a single model?

Even a single model in production benefits from basic MLOps practices, like versioning and monitoring, since models degrade over time regardless of how many are running. MLOps becomes more clearly necessary once a team is managing more than a handful of models at once.