Heterogeneous data-driven resolution generation for software systems via large language models

Abstract

Modern software systems are increasingly complex and dynamic, making them particularly vulnerable to performance anomalies. Although runtime anomaly detection enhances system reliability, engineers still devote considerable time and effort to resolving errors once anomalous logs or metrics are detected. Such challenges call for intelligent automation capable of delivering targeted remediation steps based on detected anomalies. In this work, we first construct an anomaly-related knowledge base by combining heterogeneous operational data, including logs and metrics, with resolutions annotated by domain experts. Furthermore, we propose HASolver, the first Heterogeneous Anomaly Solver to generate recommended resolutions for multi-source system anomalies. The core component is a dual-view multi-vector module, designed to represent heterogeneous anomaly chunks from different modalities and to support effective multi-vector retrieval. HASolver integrates a large language model with domain knowledge to generate mitigation resolutions. We conduct extensive experiments using BLEU and ROUGE-1/2/L metrics. Compared to baseline approaches, HASolver delivers notable performance gains, improving BLEU and ROUGE-L scores by 14.6% and 19.6%, respectively. Further analyses are carried out to explore various multi-vector configurations and the effect of prompt strategies. We also release the annotated resolution dataset derived from the anomaly-related knowledge base to facilitate future research.