EUVIMEDEUVIMED entdecken
Crossref

LLM-Based Root Cause Analysis and Automated Incident Response for Large-Scale Distributed Commerce Systems

Sohail Sayed, Nauman Sayed

IOSR Journal of Computer Engineering · 2026-05 · 28 (5) · 47-61

DOI: 10.9790/0661-2805014761 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IOSR Journal of Computer Engineering, 2026-05. Verfasst von Sohail Sayed und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Production incidents in large-scale distributed commerce systems are expensive and inevitable: one hour of downtime is estimated to cost Amazon US$100 million on major shopping days (Ahmed et al., 2023), Amazon lost US$66,240 per minute of server downtime in early 2016, and the average US data center outage cost US$740,357 per event at roughly US$7,900 per minute (Nguyễn et al., 2018). Yet incident management remains dominated by manual toil: on-call engineers spend multiple rounds of back-and-forth communication to identify root causes and mitigation steps (Ahmed et al., 2023). This paper argues that large language models are the missing reasoning layer that converts incident management from a human-bottlenecked diagnostic process into an automated, governed loop: LLMs supply the causal reasoning, knowledge bases supply the evidence, and deterministic execution supplies the safety. We propose IRIS-SC, a closed-loop framework coupling (i) multisource anomaly detection over metrics, logs, and traces, (ii) alert-type-matched, evidence-grounded candidate generation, (iii) LLM-based root cause analysis with knowledge grounding and provenance, (iv) automated incident response blending playbooks with learned recovery policies, (v) continual Ops-domain benchmarking, and (vi) guardrail governance with human-in-the-loop escalation. The framework consolidates the reported evidence envelope: an LLM-driven cloud-native observability system reduces mean time to resolution by 84.2% relative to rule-based methods, achieves 95% F1 in anomaly detection, and autonomously resolves 91% of common incidents (Wang et al., 2025); an LLM-plus-RAG self-healing architecture achieves up to 82% MTTR reduction with an 89.5% autonomous success rate (Sharma & Jain, 2026); RCACopilot automates root cause analysis on a year of Microsoft incidents with accuracy up to 0.766, its diagnostic component in production for over four years (Chen et al., 2024); a production AIOps observability framework improves MTTD by 68%, MTTR by 54%, and availability by 12% (Singh, 2025); industry syntheses report 40–60% MTTR reductions for LLMpowered incident assistants and a 30–70% reduction in incident response time across implementing organizations (Damarched, 2026); a graph-augmented LLM agentic workflow improves root-cause-identification accuracy by up to 42.22% over state-of-the-art methods (Tian et al., 2025); and OpsEval, a 7,334-question benchmark across 24 LLMs, demonstrates a 0.9185 human-agreement coefficient for Ops-domain evaluation (Liu et al., 2025). The paper argues that grounded diagnosis — not fluent generation — is what separates LLM assistants from LLM operators, and that automation rates near 90% are achievable only when every reasoning step is bound to evidence and every action passes governance.

Zitieren

APA

Sayed, S., & Sayed, N. (2026). LLM-Based Root Cause Analysis and Automated Incident Response for Large-Scale Distributed Commerce Systems. IOSR Journal of Computer Engineering, 28(5), 47-61. https://doi.org/10.9790/0661-2805014761

Vancouver

Sayed S, Sayed N. LLM-Based Root Cause Analysis and Automated Incident Response for Large-Scale Distributed Commerce Systems. IOSR Journal of Computer Engineering. 2026;28(5):47-61. doi:10.9790/0661-2805014761

BibTeX
@article{Sayed2026,
  author = {Sayed, Sohail and Sayed, Nauman},
  title = {LLM-Based Root Cause Analysis and Automated Incident Response for Large-Scale Distributed Commerce Systems},
  journal = {IOSR Journal of Computer Engineering},
  year = {2026},
  volume = {28},
  number = {5},
  pages = {47--61},
  doi = {10.9790/0661-2805014761}
}
Kostenlos starten →