Sistema de detecção de anomalias e geração de recomendações em pipelines CI/CD do GitLab sem acesso ao código-fonte

Abstract

In large-scale software development environments, Continuous Integration and Continuous Delivery (CI/CD) pipelines frequently exhibit bottlenecks, instabilities, and rework that impact productivity and increase operational costs. The systematic identification of these problems still relies on manual inspections, an approach that does not scale in organizations with hundreds of repositories. Furthermore, existing solutions generally require access to source code, which may be unfeasible in environments with rigorous security and compliance policies. This work proposes a system for automatic analysis of GitLab CI/CD pipelines that detects anomalies and generates optimization recommendations operating exclusively with execution data collected via API, without requiring access to source code or configuration files. The solution combines descriptive statistical analysis (robust percentiles, z-score) with unsupervised machine learning (Isolation Forest and K-Means), enabling the identification of anomalous patterns and the generation of explainable recommendations. The system was developed using Python, PostgreSQL, FastAPI, and Streamlit, and validated with public GitLab projects. Results demonstrate that the use of adaptive thresholds per cluster reduces false positives by 30–40% compared to fixed global thresholds. Qualitative validation with professionals in the field (𝑁 ≈ 3–5) indicated high utility of recommendations. The system identified optimization opportunities with estimated gains of 40–60% in test execution time, 15–25% reduction in failure rate, and 50–70% gain in build time. It is important to note that, as this analysis was performed on public third-party repositories, these gains represent theoretical projections based on historical data and statistical comparison, not empirical measurements of real interventions implemented. As a contribution, this work demonstrates the applicability of artificial intelligence techniques to a practical DevOps problem with privacy constraints, offering an open-source tool for continuous optimization of CI/CD pipelines without compromising security and compliance requirements.