一句话：提出Aloha框架，基于对比分析和人在环智能体，实现大规模云系统批量故障的异常定位，提升处理效率与可解释性。
大规模云系统中，批量故障指同一主题下短时间窗口内大量实例受影响，通常共享根因，其异常定位耗时且认知负担重，现有方法支持不足。本文提出Aloha，一个基于对比分析和人在环智能体的框架，操作化整个批量故障处理流水线，为工程师提供场景和数据感知的指导以及可解释的根因模式。在微软云的真实批量故障案例上进行试点，结果表明Aloha简化了数据处理，支持基于对比的异常定位，使过程更实用和可访问，是迈向以人为中心、可扩展的大规模云系统故障管理的有希望的一步。
Large-scale cloud systems underpin modern computing, hosting diverse components to deliver critical services worldwide. A single fault—such as an outage or misconfiguration—can simultaneously disrupt thousands of users. Such large-scale faults, referred to as batch failures, are characterized by many affected instances across the same subject within a short time window, typically stemming from a shared root cause. Handling these failures efficiently requires anomaly localization, but existing approaches offer insufficient support to engineers, making the process time-consuming and cognitively demanding. To address this, we propose Aloha, a human-in-the-loop agent framework for anomaly localization based on contrast analysis. Aloha operationalizes the entire batch failure handling pipeline, providing scenario- and data-aware guidance along with interpretable root-cause patterns for engineers. Pilots on real-world batch failure cases in Microsoft's cloud show that Aloha streamlines data handling, supports contrast-based anomaly localization, and makes the process more practical and accessible, offering a promising step toward human-centered, scalable failure management in large-scale cloud systems.
@inproceedings{ZhangWRSGZLLZRB26,
title = {Aloha: Localizing Batch Failures in Large-scale Cloud Systems via Contrast Analysis and Human-in-the-Loop Agent},
author = {Shenglin Zhang and Yujia Wu and Jinghuan Ren and Yongqian Sun and Wenwei Gu and Chaoyun Zhang and Liqun Li and Qingwei Lin and Dongmei Zhang and Saravan Rajmohan and Chetan Bansal and Minghua Ma},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805241}
}