A comprehensive GitHub Actions dataset for CI/CD pipeline failure forecasting
Description
The GitHub Actions CI/CD Failure Prediction Dataset (GADFPD) is a curated research dataset that is designed to facilitate machine learning research that predicts Continuous Integration and Continuous Deployment (CI/CD) pipeline failures. The dataset is made up of software repository metrics, workflow execution details, commit metadata, and historical pipeline results from Open-Source GitHub repositories that use GitHub Actions as their CI/CD platform. Every record in the dataset is a GitHub actions workflow, and is classified by the outcome of the workflow run. The data consists of attributes from software repositories and CI/CD pipelines, such commit characteristics, developer activity, repository statistics, workflow configuration, execution metadata, and pipeline history. These features are designed to help develop and test predictive models that can predict pipeline failures before they occur. The dataset was created for an academic research project titled “A Machine Learning Approach for Forecasting CI/CD Pipeline Failures Using Software Repository and Pipeline Data.” It can be used for software engineering research, DevOps analytics, repository mining, machine learning, and predictive software quality research. Potential applications include CI/CD failure prediction, build health monitoring, software quality assessment, and intelligent DevOps decision support systems, among others. The dataset is made available for research and educational purposes, and is meant to encourage reproducibility, benchmarking, and further study of data-driven methods to make modern CI/CD pipelines more reliable and efficient. It builds upon public information on GitHub Actions workflows provided by GitHub and is compatible with the growing body of studies leveraging execution histories of GitHub Actions for empirical software engineering research.
Files
Steps to reproduce
Data Collection and Reproduction Workflows This data was generated to capture a full history of the Continuous Integration/Continuous Deployment (CI/CD) pipeline runs and repository commit history. Data was compiled via a custom extraction workflow written in Python, which incorporates historical CI data and newly scraped, real-time repository data from GitHub. A. Software, Instruments, and Environment • Environment: The extraction script was executed in a Google Colab environment. • Language: Python 3.x • Key Libraries: PyGithub (interactions with GitHub), GitPython (deep cloning of repositories, extracting commit histories), pandas (data manipulation and merging), and requests (download archival datasets). • Authentication: The interactions with the GitHub API were authenticated with a Personal Access Token (PAT) to boost the rate limits for the API. B. Protocols and Workflows: Data collection followed a multi-stage automated process: 1. Archival Data Ingestion (TravisTorrent Baseline). • Protocol: A historical dataset (final-2017-01-25.csv.zip) was downloaded from Figshare (TravisTorrent), and copied to the working folder. • Processing: The processing script extracted the latest 100,000 records and mapped the legacy column names to a modern and unified data model to make the data compatible with modern GitHub Actions (GHA) data. 2. Live Pipeline Metadata Extraction (GitHub API) • Target Selection: The list of Targets to be selected is a pre-defined list of high-profile open-source repositories in multiple domains. • Protocol: For each repository, the first 25,000 recent GitHub Actions workflow runs were retrieved from the GitHub API. • Data Captured: Head_branch and event triggers were captured along with execution timestamps and final pipeline status/conclusion for each workflow run. 3. Deep Git History Extraction (Local Cloning). • Protocol: At the same time, we used the GitPython library to locally clone the target repositories up to 35,000 commits deep. • Data Captured: The script went through all commits in the git history and captured various commit-level metrics such as developer login, commit message, timestamp, total code churn (insertions/deletions), and number of modified files. 4. Data Integration and Deduplication • Merging Protocol: The commit metadata in the GitHub Actions pipeline and the deep Git commit history were joined in a pipeline using an inner merge with the commit_sha as the primary key for each mined repository. This resulted in a unified view mapping the code-level changes to pipeline execution results. In addition, the newly joined data from GitHub was concatenated with the normalized historical data from TravisTorrent. • Refinement: After a strict deduplication process, the final merged dataframe contained a highly refined and unified data set of 303,079 unique CI/CD execution records for the years 2016–2026. The output was generated as a master CSV file.
Institutions
- Babcock UniversityOyo State, Ibadan