Scale Your Machine Learning with MLOps

Editor's note: This article was originally published on our sister site, InformationWeek.

A few years ago, everyone was trying to figure out how to get started with artificial intelligence and one of its components, machine learning. But today many organizations have put together pilot programs, identified promising use cases, and even turned around some value for their organizations.

After you've won those initial successes, it's time to expand that value to other use cases and other parts of the organization. But with each of your initial use cases you learned something. You developed some technology that you may want to use again. You identified approaches that may not have worked as well as others. How do you take those lessons and apply it to new projects? How do you ensure that you are not re-inventing the wheel each time you tackle a new data science job?

Machine learning ops, or MLOps, takes the principals and philosophy behind DevOps and applies them to the practice of data science. Omdia chief analyst for AI platforms, analytics, and data management, Brad Shimmin, shared his perspective on the value of developing an MLOps practice and how to go about it during his presentation at Interop Digital, MLOps: Supercharging Data Science in the Enterprise.

"The more you invest in AI, the more you optimize, the more you instrument your business and then use AI to understand it and to predict it, and to accommodate the changes that might happen, then your business becomes more resilient to change and more agile," Shimmin said. "You can't really get that if you're just focusing on one or two spot solutions like just trying to figure out turn rates or sales quarterly numbers."

To achieve that kind of a machine learning practice, you need to solve for three problems: repeatability, scalability, and surety, Shimmin said.  

Repeatability means achieving the same results and being replicable. Scalability means you have enough processing power for the job you need to do. Surety means that you can trust the outcome and you can explain how the outcome was achieved. To solve these problems, you need more than just a couple of data scientists and a Jupyter Notebook, according to Shimmin.

"You need to have a process, a lifestyle," he said. "This is a collaborative sport. You're not just your data scientists, but you have data engineers, you have business analysts, you have executive sponsors. There are many, many different roles that play a very important part. One of the least understood and the least cared for is IT operations and development, and MLOps is a way to really help with that."

MLOps, however, isn't just combining data science with DevOps, because data science and DevOps are different in many ways. For instance, machine learning projects are iterative and exploratory, and are constantly revising themselves, Shimmin said.

"There's so much open source technology available that it makes it almost impossible to really lock down a given architecture that is going to be consistent across just two projects next to one another," he said.

Plus, models and data change over time. The moment a model goes live, it starts to degrade for those very reasons.

Because of the changing nature of a data science practice, a platform for MLOps needs to be able to accommodate for all these changing and evolving components. Shimmin offered a big list of many of the different functions that such a platform would require. They include the following: an ML project artifact repository; data and feature catalog; intrinsic system resource provisioning; flexible model deployment options; a cloud-native utility business model; open source friendliness; multi and hybrid cloud capabilities; high availability and disaster recovery; quality-specific tooling; and other features, too.

"Identifying the right features to use and then crafting them such that they can be consumed and used by an AI algorithm is really, really tough," Shimmin said. "Having a place to house and then reuse those is extremely important. You have to have the ability to provision all those resources."

The platform solution that you choose to manage MLOps must accommodate "changes in the marketplace and be able to bring in libraries, frameworks, and other tools without having to accrue any sort of hidden technical debt in managing them," Shimmin said. "That's what these platforms are all about."

Omdia is creating an Omdia Universe report expected to be released later this year that will look at MLOps pure-play platform vendors, many designing their platforms for the enterprise, and an early list includes the following: Cnvrg.io, Comet, Dataiku, dotData, Iguazio, Pachyderm, Tecton, and Verta.

Shimmin acknowledges that these platforms are not perfect and remain relatively immature. Many issues remain to be solved, such as how to move from one platform to another, how to work with data versioning, and how to provide for data security, governance and management.

"At the end of the day you've got an emerging class of MLOps platforms that because they are so open, because they are cloud native, you can take all the investments you've already made in data science and start feeding them into these platforms."