Abstract

Due to the lack of practical guidelines on how to write logging statements and large volume of logs routinely generated by software products, how to make proper logging decisions and efficiently analyze the logs are challenging in practice. In this thesis, we focus on these two main challenges and propose a series of approaches to address the problem and help developers on logging practices in two aspects: (1) assist in making logging decisions and (2) assist in log analysis.

For logging decisions, we tackle the challenge by providing suggestions on writing logging statements. We first provide suggestions for logging locations. We find that our models are effective in suggesting logging locations at the block level. We then study the verbosity levels in the logging statements. We propose a deep learning based approach that can leverage the ordinal nature of log levels to make suggestions on choosing log levels. Our approach outperforms the baseline approaches and are effective at suggesting log levels. Finally, we investigate practitioners' expectation on the readability of log messages by conducting a series of semi-structured interviews with industrial practitioners. We derive three aspects that are related to the readability of log messages. We also explore the potential of automatically classifying the readability of log messages and find that both deep learning and machine learning approaches is effective at such classifications.

For log analysis, we focus on studying log abstraction, which is a crucial step for automated log analysis. We find that different categories of dynamic variables in logs record valuable information that can be important for different tasks, such information is abstracted by prior log abstraction techniques. We propose a deep learning based log abstraction approach, which can identify different categories of dynamic variables and abstract specified categories. Our approach outperforms state-of-the-art log abstraction techniques on general log abstraction and also achieves promising results on variable-aware log abstraction. We also find that variable-aware log abstraction can help improve the performance of log-based anomaly detection.