85% of AI projects ultimately fail. Why? It’s not just that building and training models have their challenges. Addressing the end-to-end challenges of delivering production AI requires tooling developed for the specific case of DevOps of machine learning â€“ MLOps.
Despite the increased interest in and adoption of artificial intelligence (AI) in the enterprise, 85%Opens a new window of AI projects ultimately fail. This is in part because data scientists have difficulty deploying machine learning (ML) models into production due to insufficient tooling and ad-hoc ML development processes. Organizations need the right technology and processes in place when developing AI applications in order to unblock ML deployment and deliver business value from AI.
Insufficient tooling and ad-hoc processes have led to difficulty operationalizing AI in the enterprise. Because of that, data science teams struggle with deploying ML models before they make it into production. Based on a survey of 500 industry professionals, the State of Development and Operations of AI Applications 2019Opens a new window market research found that despite enterprises dedicating significant time and resources towards their AI initiatives, many data science and ML teams do not have the proper tools required to efficiently collaborate, build, scale and deploy models.
Additionally, most companies want to leverage an open, hybrid, multi-cloud approach to avoid issues including lock-in and data migration. For example, many enterprises have existing data lakes and processes, and are not cloud-native, making solutions from public cloud platforms, such as Amazon AWS, Microsoft Azure and Google Cloud Platform ill-suited. Moreover, ML platforms that claim to have end-to-end capabilities with modeling functionality are also unsuitable if the resulting data flows and models are not auditable and reproducible. Thus, organizations should consider applying DevOps practices to ML development.
DevOps has transformed the way software engineers deliver applications by making it possible to collaborate, test and deliver software continuously. However, AI and data science today is like software engineering was in the 1990s, with data science and ML teams working with broken processes that lead to difficulties with collaboration, error-prone manual tracking, no reproducibility or provenance, lack of automated testing and unmonitored models. As a result, the processes and tools for collaborating and maintaining ML projects at the industrial scale are not yet as mature as for traditional software projects when using DevOps techniques. ML engineering should be just as easy, fast and safe as modern software engineering when using DevOps techniques. This is why the same DevOps techniques that are used to address the end-to-end challenges of software development can be applied to ML development as a way to solve difficulties with AI in production today.
Furthermore, data science and machine learning contain unique requirements that are not well covered by existing DevOps tools. Large datasets need to be prepared and tracked, machine learning models have adjustable hyperparameters that must be set correctly, and a deployed model will degrade over time as incoming data changes, meaning it must be constantly monitored. This means that new tools must be developed to address the specific case of DevOps for machine learning.
It is not just startups that are facing this challenge. In fact, Fortune 100 companies are also realizing the challenges of deploying ML models into production. For example, S&P Global’s Director of Machine Learning, Ganesh Nagarathnam, understands this challenge well and is an advocate for applying DevOps principles to ML development in order to unblock AI in the enterprise. He says, â€œMore than 50% of AI models get blocked before they reach production. By applying the DevOps practices of collaboration, control and continuous delivery to AI, teams are able to reliably transition AI models from ideation, rapidly through testing ideas to get them ready for production, and finally into productization.â€
Developing AI applications requires an end-to-end approach and organizations need the right technology that enables ML and data science teams to collaborate, test and deliver software continuously â€“ ultimately delivering business value for businesses. Through the use of DevOps techniques and practices, data science and ML teams will be able to track AI runs, maintain a complete audit trail and provide total visibility into an ML model’s provenance through the AI lifecycle, collaborate more efficiently and deliver new models to production quickly while remaining in control of quality and addressing regulatory considerations.
Applying DevOps best practices to ML development removes the productivity and collaboration challenge data science and ML teams experience, enabling them to control the entire model development and operations process from idea to production.