Integrating Automated Profiling an‎d Tracing into CI/CD Pipelines for Performance Regression Detection

2/20/2026 12:00:00 AM

The quick adoption of microservice designs an‎d Continuous Integration/Continuous Deployment (CI/CD) processes has sped up the supply of software, but it has also made "silent" performance regressions a big problem. These non-functional degradations, such latency spikes an‎d resource leaks, often get past regular quality assurance checks, which can lead to system instability. This thesis offers an automated, machine learning-driven quality gate aimed to proactively discover performance irregularities inside the CI/CD pipeline prior to production deployment. This thesis employs a stringent trace-driven simulation technique an‎d the Train Ticket microservices benchmark dataset to assess three unsupervised learning algorithms: Isolation Forest, One-Class SVM, an‎d Deep Autoencoder. The study utilizes high-fidelity telemetry data, characteristic of eBPF-based distributed tracing, to simulate intricate, multivariate system behaviour without the necessity of pre-labelled failure records. The experimental findings show that the Deep Autoencoder model works better, with an F1-score of 0.8665 an‎d a ROC-AUC of 0.9383. Most importantly, the suggested framework improved detection accuracy by 46% over existing statistical baseline approaches an‎d cut down on false positives by 76%. This solved the problem of al‎e‎rt fatigue in the field. The model meets the latency requirements of modern pipelines because it can make inferences in less than 100 milliseconds. This study validates that the incorporation of unsupervised deep learning into the software delivery lifecycle offers a resilient, scalable solution for guaranteeing performance dependability while maintaining deployment speed.

معماري ميكروسرويس ها , تشخيص ناهنجاري

Microservices Architecture , Anomaly Detection

Hanan Helta

Dr. Azhari