Towards addressing MLOps pipeline challenges : practical guidelines based on a multivocal literature review

Abstract

Machine learning and artificial intelligence have been adopted in many businesses recently. The adoption of continuous software engineering practices such as DevOps in deploying machine learning projects to production is termed as MLOps. However, not all ML projects reach production due to the various complexities involved. This paper presents the challenges present in different components of the MLOps pipeline namely the data manipulation pipeline, model creation pipeline, and deployment pipeline. We conduct a systematic literature review and grey literature review to identify the challenges in the MLOps pipeline. Based on this data, we synthesize practical and relevant guidelines for practitioners, e.g., enriched with available tool support that could be used to implement them. The applicability of a selection of these guidelines is then demonstrated qualitatively by using them in the context of an example case with industry-relevant ML components. The contribution of this paper is threefold. First, we review contemporary literature to provide an overview of the state-ofthe- art in MLOps challenges, solutions, and tools used. Second, we present seven practical guidelines based on the review. Third, we apply these guidelines with practical demonstration. The results will provide insight into how certain MLOps challenges can be overcome by following guidelines (not tool specific) mentioned in our study in the area of research and industry.