An architectural perspective on MLOps: Structures, processes, tools, and stakeholders

Despite the increasing adoption of Machine Learning Operations (MLOps), teams still encounter challenges in effectively applying this paradigm to their projects. While numerous MLOps tools exist, consolidated knowledge to inform architecture design is still lacking. In response, our goal is to provide a comprehensive overview of MLOps architectures from a structural and process perspective, the tools mentioned supporting the implementation of architecture components, and the stakeholders responsible for the MLOps process. We conduct a systematic mapping study of 93 primary studies to collect and analyze the state of the art knowledge on MLOps systems using automatic, manual, and snowballing-based search strategies. Subsequently, we use card sorting to synthesize the results. We contribute: (i) a categorization of 39 MLOps architecture components and a description of several MLOps architecture variants; (ii) a systematic map between the components and the existing MLOps tools; (iii) a description of 56 process steps for MLOps systems creation, deployment, and maintenance; and (iv) a description of MLOps stakeholders, their responsibilities, and a systematic map between the process steps and the responsible stakeholders. Our results serve as an overview of the state of the art in MLOps architectures from a structural and a process perspective to support researchers and practitioners in the architecture design of their MLOps systems.