ADAHEAL: AN ADAPTIVE FRAMEWORK FOR FAULT DETECTION AND AUTOMATED RECOVERY IN LARGE-SCALE CLOUD-NATIVE MICROSERVICES
AI and Data Science Journal · 31.12.2024 · 1 (1) · 132-141
DOI: 10.61784/adsj4045 ↗Kurz gefasst
Fachartikel in der Zeitschrift AI and Data Science Journal, online erschienen am 31.12.2024. Verfasst von Xiang He und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Cloud-native microservice systems fail in ways that are frequent and hard to contain: their telemetry drifts with every deployment and traffic surge, faults propagate along call graphs and surface far from their origin, and the effectiveness of remediation actions changes as the system evolves. This paper presents AdaHeal, an adaptive closed-loop framework that unifies fault detection, root-cause localization, fault-type identification, and automated recovery. AdaHeal detects anomalies with a dual-axis masked reconstruction network (DAMR) that reconstructs every telemetry cell only from its temporal and cross-metric context and keeps alarms calibrated under benign drift through score-gated online recalibration. It localizes faults with calibrated propagation-aware ranking (CPR), which scores each service by the part of its deviation that its dependencies cannot explain under the flow-conservation rules of synchronous RPC and assigns latent evidence to stateful backends without request-level metrics. Recovery is planned by a risk-aware discounted Thompson sampling engine (RA-DTS) that selects (service, action) pairs under an attempt budget and revises its beliefs after every failed attempt. On the public SMD and PSM datasets, DAMR reaches a point-adjusted F1 of 0.893 on SMD (best of nine detectors) and the highest AUPR among the deep sequence models on PSM (0.440). On microservice benchmarks built from the Online Boutique, Sock Shop, and a 100-service Alibaba-style topology, CPR raises top-1 localization accuracy to 0.964, 0.903, and 0.922 (best baseline: 0.858, 0.694, 0.892) at 60 ms per incident for 400 services. In closed-loop experiments with shifting action effectiveness, AdaHeal shortens the mean time to recovery by 35.7% relative to a static detection–localization–runbook pipeline.
Zitieren
APA
He, X., & Zhou, L. (2024). ADAHEAL: AN ADAPTIVE FRAMEWORK FOR FAULT DETECTION AND AUTOMATED RECOVERY IN LARGE-SCALE CLOUD-NATIVE MICROSERVICES. AI and Data Science Journal, 1(1), 132-141. https://doi.org/10.61784/adsj4045
Vancouver
He X, Zhou L. ADAHEAL: AN ADAPTIVE FRAMEWORK FOR FAULT DETECTION AND AUTOMATED RECOVERY IN LARGE-SCALE CLOUD-NATIVE MICROSERVICES. AI and Data Science Journal. 2024;1(1):132-141. doi:10.61784/adsj4045
BibTeX
@article{He2024,
author = {He, Xiang and Zhou, Lin},
title = {ADAHEAL: AN ADAPTIVE FRAMEWORK FOR FAULT DETECTION AND AUTOMATED RECOVERY IN LARGE-SCALE CLOUD-NATIVE MICROSERVICES},
journal = {AI and Data Science Journal},
year = {2024},
volume = {1},
number = {1},
pages = {132--141},
doi = {10.61784/adsj4045}
}
EUVIMED