Application of kubeflow as a universal approach for the development and implementation of artificial intelligence systems

Abstract

An overview of the concept of machine learning and processes (Machine Learning and Operation, MLOps), which is a set of techniques for implementation and automatic continuous integration, as well as delivery to the production environment and model training, is made. The concept of MLOps was considered in terms of Kubeflow tools - a cloud-native open-source system running on the Kubernetes platform. The possibility of using modern MLOps solutions to improve the development processes of machine learning information systems has been studied. The results of the operation of the model in the Kubeflow arsenal have been checked using such improvement factors as speed of development, implementation of changes, reduction of time to search for problems, recovery after global interruptions, and decrease of the number of errors in the model. For practical analysis, a publicly available model was deployed in a Kubeflow cluster using the Seldon Core Serving application manifest. The conducted research showed that Kubeflow consists of a set of various open-source components that have a high level of integration with each other through the Kubernetes platform. At the same time, Kubeflow uses the Kubernetes pattern of operators for machine-learning objects extremely effectively.