Machine and Operations Learning (MLOps)

DZone 's Guide to

Machine and Operations Learning (MLOps)

Learn more about machine and operations learning (MLOps) and see benefits and the different phases.

· AI Zone ·
Free Resource

Machine and Operations Learning (MLOps), similar to DevOps, is a combination of practices and tools that use Data Science and Operations teams to improve model deployment through Machine Learning (ML), AI, monitoring, validation, collaboration, and communication.

Figure: Microsoft Azure - MLOps

This is because according to Gartner, many companies develop Machine Learning models, but only 47% of them are published in production. And still, 88% of AI initiatives have difficulty passing the test stage.

Azure ML consists of web interfaces and SDK that do the training and deploy the model. The process (figure below) involves data management, data cleansing, writing and displaying experiments, publishing models to collect actual data (in production) and applying model improvements.


  • Experiment: Data scientists conduct a variety of experiments, evolving and collecting data, seeking answers to business needs. Fundamentals of DevOps and Software Engineering concepts, or MLOps, suggest practices to avoid promoting poorly written models in other environments.
  • Develop: Training of algorithms (Azure ML Training Services), data pipelines (ETL), and how to pipette CI/CD practices (Azure DevOps pipelines) to deploy built ML experiments.
  • Operate: Inference, monitoring (using analysis/profiling, application telemetry, etc.), automated testing, and data feedback loop for learning from data leading to model improvements.
You might also want to read:  Intro to Machine Learning for Developers


With the teams working together (data scientists and operations), encouraging as data-driven business operations, using range or the discovery of insights and implementation actions. Some of the advantages are:

  • Reproducibility and Audibility: Creation of models in applicable and reproducible pipelines, enabling rollbacks (in case of errors) and audits if tracking is required.
  • Validation: Inherits DevOps concepts such as automated validations, testing, profiling, and environment management.
  • Enable automation and observability to perform new deployments. Allows a comparison of expected vs. expected performance. Collection of information that serves as model training for future improvements.

MLOps Phases

The DevOps framework for AI solutions in Azure can be macro represented as follows:

And it comprises four main phases from design creation to deployment:

  1. Model creation and training: Use learning pipeline training to create reproducible pipelines, bringing together all the steps (from data preparation to model evaluation).
  2. Model Deployment: We package the model for deployment and profiling to help configure memory, CPU, and validate models.
  3. Automate or learn E2E: Frequently update templates with Azure Machine Learning and GitHub, test and implement enhancement with other applications and services.
  4. Audit trail: Collect end-to-end data to define an audit trail. For example: model publisher data, implementation data, production use, etc.

Figure: Microsoft Azure - MLOps (adapted)

Further Reading

Understanding Machine Learning

Machine Learning: An Introduction

machine learning ,devops ,mlops ,machine and operations learning ,mlops phases

Published at DZone with permission of Leonardo Matsumota . See the original article here.

Opinions expressed by DZone contributors are their own.

{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}