Meet Us at ODSC West in San Francisco from Oct 31-Nov 1

Adopting a Production-First Approach to Enterprise AI

Yaron Haviv | December 16, 2021

After a year packed with one machine learning and data science event after another, it’s clear that there are a few different definitions of the term ‘MLOps’ floating around. One convention uses MLOps to mean the cycle of training an AI model: preparing the data, evaluating, and training the model.  This iterative or interactive model often includes AutoML capabilities, and what happens outside the scope of the trained model is not included in this definition. My preferred definition of MLOps refers to the entire data science process, from ingestion of the data to the actual live application that runs in a business environment and makes an impact at the business level. This isn’t just a semantic issue--this discrepancy in usage has major consequences for the ways enterprises bring ML to production—or don’t.  

The Training-First Mindset in MLOps 

As a vestige of an era when ML was confined mostly to academia, or to analytics services, most data science solutions and platforms today still start with a research workflow and fail to deliver when it comes time to turn the generated models into real-world AI applications. So much so, that even the term ‘CI/CD pipeline’ is sometimes used to refer to the training loop, and not extended to include the entire operational pipeline. This mindset forces the ML team to re-engineer the entire flow to fit the production environment. At this late stage, building the actual AI application, deploying it, and maintaining it in production become acutely painful, and sometimes nonviable.  

Modern applications in which AI models provide real-time recommendations, prevent fraud, predict failures and guide self-driving cars require significant engineering efforts and a new approach to make it all feasible.  

Adopting a Production-First Mindset 

I advocate for a mindset shift. Begin with the end in mind: that is, design a continuous operational pipeline with a production-first approach, and ensure that the various components and practices hook into it. To scale your organization’s AI adoption, the goal at the outset should be make the process repeatable so that you can scale along with the organization’s needs. 

To read more, including the four key components for production pipelines, read  my full piece on Towards Data Science.