ML Ops: Operationalizing Data Science

Would you spend many years and big money training athletes and then send them to the Olympic Games, only to make them stay in their hotel instead of competing?

“Of course not,” you say. “That would be ridiculous. Nobody would do that.”

You’re right. It is ridiculous.

But if you’re spending a lot of time and money developing and training your analytics and machine learning (ML) models without getting them into production—be that because of operational difficulties or because models are not consistent with applicable regulations and laws—aren’t you making the same mistake? Models that do nothing more than provide static insights in a slideshow are not truly “operational,” and they don’t drive real business change.

You’re probably not setting your models on the shelf deliberately. There are plenty of reasons why so many models—by some estimates, more than half—don’t make it into production. But if they’re not in production, they can’t do what you’ve trained them to do.

ML Operations, or ML Ops, is the process of operationalizing data science by getting ML models into production—being able to monitor their performance and ensure they are fair and in compliance with applicable regulations. The four main steps in the process (Build, Manage, Deploy and Integrate, and Monitor) form a repeatable cycle for handling models as reusable software artifacts. ML Ops ensures that models continue to deliver value to the organization, while also providing critical ...