Reproduce “A Survey of AIOps for Failure Management in the Era of Large Language Models” on OpenResearch

About this paper

Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Aiwei Liu, Yong Yang, Zhonghai Wu, Xuming Hu and 2 more

The paper provides the first comprehensive survey of Large Language Model (LLM) applications for failure management in Artificial Intelligence for IT Operations (AIOps), establishing taxonomies for data sources, LLM approaches, and AIOps tasks. It details how LLMs address limitations of traditional methods across the failure management lifecycle, from data preprocessing to auto remediation, while identifying key challenges like computational cost and hallucinations.

Key insights

• LLMs significantly enhance AIOps by effectively processing both structured system data and unstructured human-generated data, which was previously underutilized.
• LLMs improve the generalizability of tasks like log parsing, augment anomaly detection with semantic understanding, and enable holistic root cause report generation rather than isolated classifications.
• Despite their transformative potential, LLM-based AIOps faces critical challenges including high computational costs, underutilization of complex data sources (e.g., trace data), difficulties in achieving consistent generalizability in evolving systems, and the risk of hallucinations.

Summary by alphaXiv.