Análise comparativa da implantação de um modelo de classificação com aprendizado de máquina em diferentes níveis de MLOps

Abstract

The integration of machine learning (ML) application implementations with system operations is notably more complex compared to other applications in the field of Software Engineering. This is due to ML systems facing the same challenges as common software systems, compounded by the difficulties inherent in ML. Maintaining these systems is also a complex task due to the unstable nature of ML models when confronted with variations in data patterns.. This work conducts a comparative analysis between different levels of MLOps to understand the nuances of this integration. Adopting an experimental comparative methodology, we conducted a literature review on MLOps and set up a working environment on Amazon Web Services (AWS). We implemented a classification system at three different MLOps levels, assessing efficiency, quality, and ease of maintenance. The results reveal that tools like MLFlow for metric tracking and logging, along with solutions like Evidently for data monitoring in production, play crucial roles in creating a robust and reliable ML ecosystem. This study provides valuable insights into the differences between MLOps levels, emphasizing the importance of these methodologies in the broader context of integrating machine learning systems.