一句话:用 DeepSeek-V2 与 GPT-4o 大规模解析 28,770 个 GitHub 仓库的 CI/CD 流水线定义,量化各类实践的采纳率与系统差异。
CI/CD 流水线配置是理解现代软件开发实践的重要窗口,但面对数量庞大、写法各异的流水线定义文件,传统方法难以规模化地解析与归类。本文利用 DeepSeek-V2 和 GPT-4o 两个大语言模型,对 28,770 个流行 GitHub 仓库的流水线定义文件进行解析与语义解读,统计各类 CI/CD 实践的采纳趋势,并比较不同 CI 系统的侧重。结果显示,构建(80.0%)与测试(80.2%)阶段近乎普及,而静态应用安全测试 SAST(12.1%)、容器化(17.9%)和云部署(4.2%)等专门实践的采用率明显偏低;系统层面差异显著,Travis CI 在构建自动化上领先(99.2%),GitHub Actions 则在 linting(44.9%)与 SAST(16.5%)上表现突出。研究不仅刻画了 CI/CD 实践全景,也展示了 LLM 在软件工程实证研究中的潜力,并提出改进 LLM 流水线分析工具、弥合采用差距的未来方向。
The analysis of continuous integration and continuous deployment (CI/CD) pipeline configurations offers critical insights into modern software development practices. This paper leverages advanced large language models (LLMs) — DeepSeek-V2 and GPT-4o — to examine pipeline definition files from 28,770 popular GitHub repositories, uncovering adoption trends across CI/CD practices. Our LLM-driven approach reveals near-universal implementation of build (80.0%) and test (80.2%) stages, contrasted by lower adoption of specialized practices like static application security testing (SAST) (12.1%), containerization (17.9%), and cloud deployment (4.2%). System-specific patterns emerge, with Travis CI leading in build automation (99.2%) and GitHub Actions excelling in linting (44.9%) and SAST (16.5%), reflecting distinct system strengths. By harnessing LLMs to parse and interpret complex pipeline files at scale, this study not only maps the CI/CD landscape but also demonstrates the transformative potential of artificial intelligence in software engineering research. We propose future work to refine LLM-based tools for pipeline analysis and address adoption gaps, aiming to enhance CI/CD accessibility and effectiveness across diverse projects.