A Theoretical Framework for Understanding the Relationship between Log Parsing and Anomaly Detection
Donghwan Shin, Zanis Ali Khan, Domenico Bianculli and Lionel Briand
Abstract: Log-based anomaly detection identifies systems’ anomalous behaviors by analyzing system runtime information recorded in logs. While many approaches have been proposed, all of them have in common an essential pre-processing step called log parsing. This step is needed because automated log analysis requires structured input logs, whereas original logs contain semi-structured text printed by logging statements. Log parsing bridges this gap by converting the original logs into structured input logs fit for anomaly detection.
Despite the intrinsic dependency between log parsing and anomaly detection, no existing work has investigated the impact of the “quality” of log parsing results on anomaly detection. In particular, the concept of “ideal” log parsing results with respect to anomaly detection has not been formalized yet. This makes it difficult to determine, upon obtaining inaccurate results from anomaly detection, if (and why) the root cause for such results lies in the log parsing step.
In this short paper, we lay the theoretical foundations for defining the concept of "ideal” log parsing results for anomaly detection. Based on these foundations, we discuss practical implications regarding the identification and localization of root causes, when dealing with inaccurate anomaly detection, and the identification of irrelevant log messages.