Should MLOps Be Verticalized?

Video produced by Steve Nathans-Kelly

In a presentation at Data Summit Connect 2021, Google AI/ML Ops specialist Skander Hannachi made the case against MLOps verticalization. Hannachi recalled watching a talk given by Andrew Ng on MLOps. In the talk, called "From model-centric to data-centric AI," Ng opined  that MLOps will eventually become verticalized. "What does that mean? That means MLOps will not remain a unified discipline applicable across multiple industries and specialties; that there'll be a special MLOps for inventory and demand forecasting. There'll be a different, completely different type of MLOps for fraud detection. And there'll be yet another MLOps for people who are working in the medical industry and it will break up into a whole bunch of separate disciplines."

Hannachi said this idea hit close to home for him because he had realized recently that a lot of the work that was the day-in, day-out of an engineering team that's responsible for a retail merchandising and demand forecasting pipeline was effectively MLOps, but  it was not called that. It was called DevOps or MerchOps. "And the reason we did so was that the tool set was radically different. It was based on highly specialized demand, forecasting, and inventory planning tools." Typically, they were proprietary and not open source and typically they were on-prem and not cloud-based. the terminology and the jargon was very different, but the mindset was exactly the mindset that an MLOps team should have, he noted.

"And so the fact that we didn't call it MLOps because it used such a highly specific technical stack, with specific terminology, and highly specific jargon was the reason why people didn't think of it as MLOps."

However, said Hannachi, he questioned whether we should be thinking about it as MLOps at all, given just how specific that tech stack and that terminology and those considerations are. "And that's why Ng's comment hit  really close to home because maybe he's right. Maybe we're never going to be able to have one general MLOps discipline across everything, but then I think about it, and no, we shouldn't limit ourselves."

According to Hannachi, it might be efficient in the short term to just leave demand forecasting teams doing it their way and allow each discipline evolve on its own and become verticalized, but we would be a lot poorer for that because people won't be able to cross-learn and skills won't be transferable. "There might be things that our fraud detection team could teach a demand forecasting team. There might be things that a fraud detection team could teach a  team that's working on applying ML to the medical field—the regulatory and ethical concerns that a financial or a mortgage-related MLOps team knows about could be very well applicable to demand forecasting or to inventory. Why would demand forecasting have any ethical considerations? Well, it might, and we can learn a lot from the people who already have to take into account ethics considerations because they're working in finance or in real estate or in fraud. So I think, in answer to Andrew Ng's comment that MLOps will eventually become more verticalized, that might be a good idea in the short term—and the long term, I don't think so and it shouldn't. But predictions about the future, as somebody who works in in data science and demand forecasting, are always tenuous. So we'll wait and see."