UC San Diego

As containerized development has grown in popularity, it’s been increasingly employed in domains for which it wasn’t originally designed, such as AI/ML. Anecdotally, it has been reported that development teams are concerned with the challenges of containerization, despite the promised gains in productivity. There have been few studies focused on understanding developer perspectives in the context of the changes to the software development cycle that arise from employing containers. Furthermore, past studies do not relate the various types of workloads to job roles and the extent of their impact on the specific development tasks performed and tools being employed in containerized development. To address this gap, we conducted 35 contextual interviews with Kubernetes cluster operators and varying levels of users (i.e., namespace users and managers). Participants were recruited from two distinct research organizations, one industrial and the other academic. Their primary focus was to train AI/ML models or use them in containerized application development. We found that cluster operators and users frequently perform a range of tasks that diverge from their job role expertise. As a consequence, they frequently switch among a multiplicity of tasks, using various tools. This drives them to create custom tool integrations to mitigate the cognitive load of switching between tools (and hence switching between tasks). In addition to a deeper understanding of AI/ML containerized application developer perspectives, our study contributes a set of development process recommendations and tooling implications, in view of prior work on tool interventions.