MLPacker: A Unified Software Tool for Packaging and Deploying Atomic and Distributed Analytic Pipelines

Abstract

In the last years, MLOps (Machine Learning Operations) paradigm is attracting the attention from the community, extrapolating the DevOps (Development and Operations) paradigm to the artificial intelligence (AI) development life-cycle. In this area, some challenges must be addressed to successfully deliver solutions since there are specific nuances when dealing with AI operationalization such as the model packaging or monitoring. Fortunately, interesting and helpful approaches, both from the research community and industry have emerged. However, further research is still necessary to fulfil key gaps. This paper presents a tool, MLPacker, for addressing some of them. Concretely, this tool provides mechanisms to package and deploy analytic pipelines both in REST APIs and in streaming mode. In addition, the analytic pipelines can be deployed atomically (i.e., the whole pipeline in the same machine) or in a distributed fashion (i.e., deploying each stage of the pipeline in distinct machines). In this way, users can take advantage from the cloud continuum paradigm considering edge-fog-cloud computing layers. Finally, the tool is decoupled from the training stage to avoid data scientists the integration of blocks of code in their experiments for the operationalization. Besides the package mode (REST API or streaming), the tool can be configured to perform the deployments in local or in remote machines and by using or not containers. For this aim, this paper describes the gaps this tool addresses, the detailed components and flows supported, as well as an scenario with three different case studies to better explain the research conducted.