This paper points out that as AI technology advances and ML models are deployed in real-world systems, the heterogeneous large data and efficient response requirements of real-world environments reveal the limitations of existing software architectures. Focusing on data-centric architecture (DOA), which has emerged as a new architecture to address these issues, we investigate how and to what extent DOA has been implicitly adopted in the actual ML-based system implementation and deployment process. Using a systematic and semi-automated methodology in software engineering, we review the papers and demonstrate that the adoption of DOA contributes to meeting requirements such as big data management, low-latency processing, resource management, security, and privacy protection, and provide practical advice for deploying ML-based systems.
