On the Challenges of Implementing MLOps for Stream Learning Algorithms

Miguel G. Rodrigues, Eduardo K. Viegas, Fabrício Enembreck, Altair O. Santin, Juliano S. Langaro, Adilson G. Filho

Abstract: Machine Learning Operations (MLOps) are essential for the efficient management of the Machine Learning (ML) lifecycle, ensuring scalability, adaptability, and operational performance across deployments. However, there is limited implementation of MLOps architectures specifically designed for stream learning scenarios, which involve continuous data flows and frequent model updates. This paper provides a comprehensive overview of existing MLOps architectures with a focus on their support for streaming data processing, model versioning, and real-time updates. Our analysis identifies critical gaps related to scalability, adaptability, and efficient handling of frequent updates in dynamic stream learning environments. We highlight the open gaps to pave the way for architectures that provide real-time model versioning and update mechanisms to improve stream learning performance across diverse application domains.

External IDs:dblp:conf/aina/RodriguesVESLF25