An Intelligent Machine Learning Framework for Early Detection of CI/CD Pipeline Failures

Abstract

Continuous integration and continuous delivery (CI/CD) pipelines automate software build, testing, and deployment workflows, but failures can delay releases and consume engineering resources. This study investigates early failure detection using an adaptive ensemble of XGBoost, bidirectional long short-term memory (Bi-LSTM), and a BERT encoder, combined through a gating multilayer perceptron. The evaluation covers 194,400 pipeline executions from Jenkins, GitHub Actions, GitLab CI/CD, Azure DevOps, and CircleCI. A candidate space of 167 features captures build metrics, log semantics, resource utilization, test outcomes, code changes, and dependencies. Nine models are compared using precision, recall, F1-score, AUC-ROC, and training time. The ensemble achieves an F1-score of 0.940 and an AUC-ROC of 0.974, compared with 0.924 and 0.961 for BERT. At checkpoints corresponding to 10%, 25%, and 50% of estimated pipeline duration, failure-detection rates are 54.2%, 78.6%, and 91.3%, with mean warning lead times of 14.1, 8.7, and 4.3 minutes, respectively. The results indicate a trade-off between earlier warnings and detection coverage. Build-duration deviation, error-keyword frequency, and CPU spikes receive the highest feature-importance scores. Interpretation is limited by unequal model feature sets, the absence of leave-one-platform-out results, and incomplete end-to-end training-cost accounting. Future evaluation should examine threshold calibration, temporal drift, and false-alarm costs before operational deployment.