一句话：总结四轮MLOps课程教学实践，提炼有效方法、学生反馈与挑战，并提出普及MLOps教育的策略。
本文关注机器学习运维（MLOps）技能缺口：部署、监控和维护生产环境中的机器学习模型通常依赖在职摸索与团队指导，而相关课程仍较新。作者以旧金山大学数据科学研究生项目四轮MLOps课程为对象，描述并比较不同教学路径，包括课程设计、工具与项目安排，总结教师认为有效的做法，汇总学生反馈，并讨论实施中的挑战。评估主要基于多轮教学反思与学生意见，而非对照实验。最后，文章提出面向本科与研究生教育者与学生的策略与机会，以降低MLOps教学门槛、提升可及性。
Machine learning operations (MLOps) is an important aspect of the machine learning lifecycle, requiring specialized skills to deploy, monitor, and maintain machine learning models in production environments. Historically, these skills were acquired primarily through on-the-job experience, often involving trial and error or mentorship within mature MLOps teams. In recent years new courses have emerged to address this skills gap. This paper explores the pedagogic approaches taken over four iterations of the MLOps course offered in the data science graduate program at the University of San Francisco. We describe the different approaches, discuss what we perceived as effective, summarize student feedback, highlight the challenges encountered, and propose strategies and opportunities to make MLOps education more accessible for both educators and students at the undergraduate and graduate level.
@inproceedings{paperbot11995,
title = {But Have You Ever Deployed a Model to Production? Experiences with Teaching Machine Learning Operations in a Data Science Curriculum},
author = {Robert Alan Clements and Nicolas Thiébaut},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3786580.3786950}
}