Studying and Complementing the Use of Identifiers in Logs

Logs contain a large amount of curated run-time information about the process of a software. Modern software systems have become more complex and larger in scale. They are typically executed in parallel or distributively, resulting in interleaved software logs and making log analysis challenging. Despite extensive research on automated logging analysis, none to our knowledge focuses on the use of logs, and they rarely augment logs to help with simpler analysis. Software log IDs are unique identifiers that developers can use to group and filter log entries. However, we found that, on average, only 21% of logging statements produce IDs, which can lead to loss of information in the log file. We propose LTID, a static analysis approach on log IDs, to remediate the aforementioned issue by extracting a dependency relation between log statements from source code. We build a dependency graph using static analysis and compute the dominance relations of each logging statement. We then propagate IDs to logs that do not contain them based on the dependency graph. We studied 21 well-known Java open-source software subjects and were able to inject IDs on average into 12% of logs without IDs. Through an open coding process, we also establish a categorization, which has a Cohen’s Kappa agreement coefficient of 0.74, of the information gained to better understand the relations recovered by the ID propagation process.