Reliable Cloud Operations Using Transformers

Managing large-scale public clouds presents significant challenges, particularly in analyzing the risk of commands executed by operators to guarantee they do not cause damages to the infrastructure. Traditional rule-based systems have been used but fall short in scalability when managing operations at a global scale. While machine learning techniques have been proposed as alternatives, they have typically been applied in small-scale environments. This article presents a novel approach to address the complexities of global-scale command risk analysis and standard operation procedure (SOP) verification using Large Language Models (LLMs). Through an extensive evaluation in production, the technique shows to be more suitable than existing methods.