CrashChecker: A Fusion Method for Clustering Duplicate Crash Failures in SAP HANA Delivery
Yang Xu, Yong Li, Qiaoluan Xie, Xiaoxiao Zhang, Chao Liu, Thomas Bach, Sunghun Kim, Sanghun Kang · 2024
To ensure high-quality SAP HANA, we thoroughly test every code change. However, a substantial portion of crash failures is duplicated during testing, and clustering these requires significant time and expertise. We propose CrashChecker, a fusion method to identify and cluster duplicate crashes efficiently. Our method assesses the probability of functions being the root cause and reranks the call stack accordingly. We then develop a trainable similarity mathematical model using our SAP HANA expertise. Additionally, we transform the call stack into a natural language format and build a deep learning similarity model by integrating a large language embedding model and deep learning modules, capturing complex representations within call stacks. Our method can effectively cluster duplicate crashes by fusing the similarity from the deep model with the mathematical model. Experiments on SAP HANA and public datasets show our method outperforms others, enabling automatic crash handling to increase testing efficiency.