par Kontaxakis, Antonios 
;Sacharidis, Dimitris;Simitsis, Alkis;Abelló, Alberto;Nadal, Sergi
Référence Data Engineering for Data Science, Springer Nature, page (141-164)
Publication Publié, 2026-01
;Sacharidis, Dimitris;Simitsis, Alkis;Abelló, Alberto;Nadal, SergiRéférence Data Engineering for Data Science, Springer Nature, page (141-164)
Publication Publié, 2026-01
Partie d'ouvrage collectif
| Résumé : | Applications based on machine learning (ML) have revolutionized many domains, including economics, health care, security, etc., which includes an ever-increasing need for fast and efficient development of pipelines. This is a trial-and-error process that involves constantly creating, executing, and evaluating new pipelines until an acceptable level of model performance is achieved. Recently, indus-try and research have developed MLOps systems to support data scientists in these tasks. These systems assist by automating pipeline design, providing execution envi-ronments, and tracking valuable metadata to support assessment. However, despite recent advancements, most MLOps systems only support parts of the design, exe-cution, and evaluation cycle, while recent surveys raise concerns about their sus-tainability and cost-efficiency. In this chapter, we first describe the main challenges MLOps systems must address at different stages of ML pipeline development. We then outline techniques used and state-of-the-art systems that implement these tech-niques to tackle these challenges. Lastly, we highlight the remaining challenges in the field and present future research directions. |
