MLOps is a very recent approach aimed at reducing the time to get a Machine Learning model in production; this methodology inherits its main features from DevOps and applies them to Machine Learning, by adding more features specific for Data Analysis. This thesis, which is the result of the internship at Data Reply, is aimed at studying this new approach and exploring different tools to build an MLOps architecture; another goal is to use these tools to implement an MLOps architecture (by using preferably Open Source software). This study provides a deep analysis of MLOps features, also compared to DevOps; furthermore, an in- depth survey on the tools, available in the market to build an MLOps architecture, is offered by focusing on Open Source tools. The reference architecture, designed adopting an exploratory approach, is implemented through MLFlow, Kubeflow, BentoML and deployed by using Google Cloud Platform; furthermore, the archi- tecture is compared to different use cases of companies that have recently started adopting MLOps.
MLOps is rapidly evolving and maturing, for these reasons many companies are starting to adopt this methodology. Based on the study conducted with this thesis, companies dealing with Machine Learning should consider adopting MLOps. This thesis can be a starting point to explore MLOps both theoretically and practically (also by relying on the implemented reference architecture and its code).