Skip to content
Nubent

/Glossary · AI

What is MLOps?

MLOps is the set of practices that applies DevOps principles to the machine learning lifecycle, from data preparation and training to deploying and monitoring models in production. The goal is models that are reproducible, auditable and easy to update.

How MLOps works

An MLOps workflow automates the steps that get done by hand in a notebook: extracting and validating data, training, evaluating against defined metrics, registering the model with its data and code versions, and deploying it if it clears the thresholds. In production, quality is monitored and retraining is triggered when the data shifts.

On AWS, Amazon SageMaker AI covers most of the lifecycle with SageMaker Pipelines for orchestration, Model Registry to version and approve models, real-time or batch inference endpoints, and Model Monitor to detect data drift. Data usually lives in S3 and is processed with Glue or EMR.

When to adopt MLOps

When you have models that drive business decisions, such as risk scoring, recommendations or fraud detection, and they are updated more than once a year. If every retraining depends on one person and their notebook, operational risk is high and reproducibility is nonexistent.

A first experimental model does not need a full platform yet. Start by versioning data and code, automating training and logging metrics. As models and teams grow, add a model registry, manual approvals and monitoring.

Best practices and common mistakes

Use separate accounts for experimentation, staging and production, and promote models between them with recorded approvals. Version training data alongside the model, define business metrics on top of technical ones, and monitor both data drift and prediction quality.

Common mistakes: deploying models without knowing what data they were trained on, skipping post-deployment monitoring and finding out months later that the model lost accuracy, and leaving oversized inference endpoints running all day when the workload is batch.

/FAQ

Questions about MLOps

Another question? Write to hello@nubent.com and an engineer will answer.

What is data drift in MLOps?

It is a change in the distribution of production data compared to the training data. It makes the model lose accuracy even though the code has not changed. You detect it by comparing input and output statistics against a baseline, for example with SageMaker Model Monitor.

Which AWS services are used for MLOps?

Mainly Amazon SageMaker AI with Pipelines, Model Registry, inference endpoints and Model Monitor, plus S3 for data, AWS Glue for preparation, Step Functions for additional orchestration and CloudWatch for operational endpoint metrics.

Does MLOps apply to generative AI?

Some of its principles do, like versioning, evaluation and monitoring. But language model applications have their own needs, such as prompt management, evaluating generated text and controlling cost per token, which fall under the term LLMOps.

Next step

What should we build next on AWS?

30 minutes with an engineer. Concrete next steps, whether or not you work with us.