A Multivocal Review of MLOps Practices, Challenges and Open Issues

随着机器学习（ML）软件应用的日益增多，ML运营（MLOps）范式引起了研究人员和实践者的极大关注。MLOps包括流程和技术，用于简化操作ML模型的资源和监控需求。软件开发从业者需要了解MLOps工作流程、实践、挑战和解决方案的详细且易于理解的知识，以有效地支持MLOps的采用。虽然学术和行业文献对MLOps的介绍正在快速增长，但相对而言，对现有大量MLOps文献进行系统综合和分析的尝试还比较少，以提高易于访问和理解的程度。我们对150个相关的学术研究和48个灰色文献进行了多声音文献综述（MLR），以提供关于MLOps的全面知识体系。通过这个MLR，我们确定了涉及各个领域的新兴MLOps实践、采用挑战和解决方案，包括开发和操作复杂管道、管理大规模生产、管理工件以及确保质量、安全、治理和道德方面。我们还通过MLOps生命周期报告了与多种角色相关的MLOps社会技术方面和跨角色的协作实践。我们断言，这个MLR为寻求在MLOps快速发展的领域中进行导航的研究人员和实践者提供了有价值的见解。我们还确定了需要解决的开放问题，以推进当前MLOps技术的发展。

MLOps practices, challenges, and solutions need to be systematically synthesized and analyzed for improving ease of access and understanding.

A Multivocal Literature Review (MLR) of 150 academic studies and 48 gray literature is conducted to provide a comprehensive body of knowledge on MLOps, including emerging practices, adoption challenges and solutions, and socio-technical aspects.

The MLR identifies various areas of MLOps, such as development and operation of complex pipelines, managing production at scale, managing artifacts, and ensuring quality, security, governance, and ethical aspects. The socio-technical aspect of MLOps is also discussed, including diverse roles involved and collaboration practices across them through the MLOps lifecycle.

Recent related studies include 'A survey on MLOps: What is it, why it matters, challenges, and future directions' and 'MLOps: Continuous delivery and automation pipelines in machine learning'.