TraceSim: An Alignment Method for Computing Stack Trace Similarity

Irving Muller Rodrigues

doi:10.1007/s10664-021-10070-w

2022-03-01

Irving Muller Rodrigues *

Eraldo Rezende Fernandes

George Chernishev

Dmitry Luciv

Nikita Povarov

DOI：10.1007/s10664-021-10070-w

原文链接

原文求助

摘要

En 中文

Software systems can automatically submit crash reports to a repository for investigation when program failures occur. A significant portion of these crash reports are duplicate, i.e., they are caused by the same software issue. Therefore, if the volume of submitted reports is very large, automatic grouping of duplicate crash reports can significantly ease and speed up analysis of software failures. This task is known as crash report deduplication. Given a huge volume of incoming reports, increasing quality of deduplication is an important task. The majority of studies address it via information retrieval or sequence matching methods based on the similarity of stack traces from two crash reports. While information retrieval methods disregard the position of a frame in a stack trace, the existing works based on sequence matching algorithms do not fully consider subroutine global frequency and unmatched frames. Besides, due to data distribution differences among software projects, parameters that are learned using machine learning algorithms are necessary to provide more flexibility to the methods. In this paper, we propose TraceSim - an approach for crash report deduplication which combines TF-IDF, optimum global alignment, and machine learning (ML) in a novel way. Moreover, we propose a new evaluation methodology for this task that is more comprehensive and robust than previously used evaluation approaches. TraceSim significantly outperforms seven baselines and state-of-the-art methods in the majority of the scenarios. It is the only approach that achieves competitive results on all datasets regarding all considered metrics. Moreover, we conduct an extensive ablation study that demonstrates the importance of each TraceSim's element to its final performance and robustness. Finally, we provide the source code for all considered methods and evaluation methodology as well as the created datasets.

Keyword:

Duplicate crash report

Crash report deduplication

Duplicate crash report detection

Automatic crash reporting

Stack trace