[论文解读] MLOps: A Step Forward to Enterprise Machine Learning

Aliza Tabassam|arXiv (Cornell University)|May 27, 2023

Big Data and Business Intelligence被引用 8

一句话总结

对MLOps及其好处、挑战和关键技术的全面评述，以及用于部署基于TensorFlow的对象检测服务的端到端企业示例。

ABSTRACT

Machine Learning Operations (MLOps) is becoming a highly crucial part of businesses looking to capitalize on the benefits of AI and ML models. This research presents a detailed review of MLOps, its benefits, difficulties, evolutions, and important underlying technologies such as MLOps frameworks, Docker, GitHub actions, and Kubernetes. The MLOps workflow, which includes model design, deployment, and operations, is explained in detail along with the various tools necessary for both model and data exploration and deployment. This article also puts light on the end-to-end production of ML projects using various maturity levels of automated pipelines, with the least at no automation at all and the highest with complete CI/CD and CT capabilities. Furthermore, a detailed example of an enterprise-level MLOps project for an object detection service is used to explain the workflow of the technology in a real-world scenario. For this purpose, a web application hosting a pre-trained model from TensorFlow 2 Model Zoo is packaged and deployed to the internet making sure that the system is scalable, reliable, and optimized for deployment at an enterprise level.

研究动机与目标

• 推动企业充分发挥AI/ML的MLOps重要性。
• 总结MLOps的演变、益处与挑战。
• 识别在MLOps中使用的核心技术、工具和框架。
• 描述在不同成熟度水平下通过自动化管道实现端到端AI项目投入生产。

提出的方法

• 对MLOps概念、工作流及生命周期阶段（设计、部署、运维）进行调研。
• 讨论底层技术与工具，如MLOps框架、Docker、GitHub Actions和Kubernetes。
• 通过从低自动化到高自动化（CI/CD和CT）的进展，说明ML项目的端到端投入生产。
• 提供一个涉及打包和部署预训练的TensorFlow 2 Model Zoo模型用于对象检测的具体企业示例。
• 描述可扩展性、可靠性和企业部署的考虑因素。

实验结果

研究问题

• RQ1 在企业环境中采用MLOps的好处与困难是什么？
• RQ2 支撑MLOps工作流的关键技术与框架有哪些？
• RQ3 MLOps如何实现具有不同自动化水平的端到端生产管道？
• RQ4 用于对象检测服务的现实世界企业MLOps部署是什么样子的？

主要发现

• MLOps框架、Docker、GitHub Actions和Kubernetes是现代MLOps生态系统的核心。
• 通过自动化管道可实现端到端ML投入生产，自动化水平从最低到最高（最小化自动化到完整CI/CD和CT）。
• 一个企业级对象检测项目演示了可扩展且可靠的部署工作流。
• 通过互联网托管的Web应用程序，可以使打包和部署预训练的TensorFlow 2 Model Zoo模型变得可访问。