A Comprehensive Review on Cloud-Native MLOps Pipelines: Tools, Architectures, and Challenges

A Comprehensive Review on Cloud-Native MLOps Pipelines: Tools, Architectures, and Challenges

Ayush Jadhav

Department  of AIML, ISBM college of Engineering, Pune

ISBM college of Engineering, Nande, Pune

Pune, India

ayushjadhavy15@gmail.com

Samar Navghare

Department  of AIML, ISBM college of Engineering, Pune

     ISBM college of Engineering, Nande, Pune

Pune, India

sam.smr918@gmail.com

Abstract - As machine learning becomes increasingly embedded in real-world applications, the need for reliable, scalable, and maintainable deployment practices has never been greater. Machine Learning Operations (MLOps) has emerged as a key discipline to address this demand, streamlining the end-to-end lifecycle of ML models. With the growing complexity of ML workflows and the shift toward distributed, cloud-based infrastructure, cloud-native MLOps has become a powerful approach. By combining containerization, microservices, and automated orchestration, it enables faster experimentation, robust deployment, and efficient model monitoring.

This review explores the landscape of cloud-native MLOps tools and architectures, focusing on widely adopted platforms like MLflow, Kubeflow, Airflow, and DVC. We categorize these tools based on their functionality, flexibility, scalability, and cloud integration. The paper also examines common architectural patterns in modern MLOps pipelines and discusses persistent challenges such as reproducibility, data drift, and system observability. Finally, we look ahead to emerging trends, including serverless MLOps, LLMOps (Large Language Model Operations), and auto-generated pipelines. This work aims to support practitioners and researchers in selecting and designing cloud-native MLOps solutions that align with their technical and organizational needs.

Keywords – MLOps, Cloud-Native, MLflow, Kubeflow, CI/CD, Data Versioning, AWS, Pipeline Automation