Detecting Cascading Failures in Microservice Telemetry: A Benchmark for Early-Warning Anomaly Detection

Cloud-native microservices-based architectures create intricate runtime dependencies and localized faults can quickly escalate to cascading service failures, making it all the more important to detect anomalies in time to ensure operational resilience. The current evaluation research mainly focuses on classification accuracy after failure and does not give much information about the early-warning ability of anomaly detection techniques in progressive failure. This paper introduces a repeatable benchmarking framework for the systematic assessment of early-warning anomaly detection, which allows to standardize the acquisition of telemetry data, to introduce controlled faults, and to employ an assessment methodology that is targeted towards the lead-time. The proposed framework has the infrastructure metrics, distributed traces, application logs, and service dependency information in a single evaluation pipeline, allowing for meaningful comparisons between the different statistical detector families, machine learning detector families, deep learning detector families, and dependency-aware detector families under the same experimental conditions. Experimental results show that dependency-aware detectors have the best overall results with an accuracy of 98.4%, a precision of 98.1%, a recall of 97.8%, an F1-score of 97.9%, an ROC-AUC of 0.97, an average lead time of 94.6 s, and the smallest delay in detection of 32.1 s. This benchmark offers a practical and reproducible way to assess and evaluate, in addition to traditional performance metrics, operational early-warning assessment, which contributes to a more resilient cloud-native monitoring and incident management system.