Why Open Source Is a Catalyst for Effective MLOps

Open source software has been behind some of the biggest technological and procedural innovations of the past decade. Its customizability and interoperability are ideal for connecting different technologies and teams and makes it an “engine of innovation.” Indeed, it’s hard to imagine the cloud, developer operations (DevOps), or containers emerging from a world without open source.

It’s also tough to envision a relatively new practice like machine learning operations (MLOps) reaching its full potential without the flexibility of open source. MLOps supports the development of intelligent applications by bringing together developers, operations managers, and data scientists in a common production environment. MLOps processes enable stakeholders to collaborate with each other to produce applications at lower cost and greater speed through a lateral flow of communication, continuous collaboration, and resource sharing.
At least, that’s the theory.

The reality for many companies has been a little less sanguine. For them, implementing MLOps has been an exercise in frustration, thanks to the use of proprietary tools or in-house solutions that are costly to build and maintain and do not integrate well with each other.

Let’s look at some of the challenges organizations are facing in their MLOps environments, how deploying an open source platform can alleviate those challenges, and how open source can be a catalyst for effective MLOps.

Bringing Different Tools and People Together

Developers, operations managers, and data scientists use different tools to do their respective jobs. For example, a developer may use a particular coding software, an operations manager might use their preferred infrastructure-as-code technology, and a data scientist might use a deep learning library for big data analysis.

Lack of interoperability between the technologies can inhibit the open exchange of information that is necessary for a successful MLOps practice. That can undermine the ability for MLOps teams to work together, adversely affecting the successful deployment of AI-driven applications.

Open source platforms can help circumvent these challenges. They support the use and integration of various tools and make it easier for different groups to share information and collaborate. For example, data scientists can use Pytorch or Tensorflow, and developers can select Argo CD or Git -- all across a common infrastructure that supports a more collaborative approach to creating and deploying models.

Accelerating Development of AI-based Applications

This kind of MLOps “virtual co-working space” doesn’t just provide stakeholders with access to tools. By eliminating the disconnect between MLOps teams, it can accelerate the development of AI-based applications.

Data scientists are becoming increasingly responsible for integrating their models into applications and monitoring their performance, so they need direct ties to developers. Likewise, developers need a direct pipeline to their data science colleagues to optimize the models for deployment and make adjustments if the models underperform. Working on an open source platform breaks down the barriers between developers and data scientists and helps them to collaborate more easily and accelerate their respective efforts.

This makes for a more productive work environment and helps organizations meet their application development and deployment goals. Companies must accelerate application development to remain competitive. Open source technology provides a good foundation for this effort by giving all MLOps stakeholders a sandbox in which to build, test, and develop at speed.

Achieving Efficiency and Addressing Concerns Around AI

What about those companies that are still on the fence about building an MLOps team, or are just getting going? According to a Deloitte survey, many companies consider themselves “starters” when it comes to AI, so it’s likely there aren’t a lot of MLOps teams out there (yet).

By democratizing AI technology, open source can help serve as an entry point for organizations and individuals just getting comfortable with AI. Even non-experts can use open source tools to gain efficiencies and develop intelligent applications. For instance, they can create custom libraries by choosing, cloning, and curating images and automatically distributing them to all data scientists. Open source projects like Thoth can help automatically optimize these libraries, as well as runtime and models, with minimal human intervention, leading to more effective applications.

It's Open Source -- Again

We’ve seen this happen before. A hot technology or process comes onto the scene and its widespread adoption is driven by interoperability, innovation, and ingenuity -- all hallmarks of the open source development community.

Here we are again. As organizations begin building their own MLOps teams, open source is leading the way. It’s bringing people closer together, giving them the tools they need, and providing the octane needed to accelerate their data science and development projects.

Will McGrath is product marketing manager, cloud storage and data services business unit, at Red Hat.