VarParser: Revolutionizing Log Parsing with Variable-Centric Techniques
Researchers from Peking University have introduced VarParser, a groundbreaking log parsing strategy that leverages large language models (LLMs) to enhance the analysis of log data. Traditional LLM-based parsers concentrate on constant log components, often neglecting valuable variable information. VarParser addresses this by employing variable contribution sampling, a variable-centric parsing cache, and adaptive variable-aware in-context learning, resulting in superior log grouping and reduced LLM invocation costs.
With extensive evaluations on large-scale datasets, VarParser demonstrates significant enhancements in accuracy and efficiency, preserving critical system information often lost in conventional methods. By focusing on variable units, the approach ensures comprehensive log analysis, vital for anomaly detection and failure diagnosis in modern online systems.
This innovative strategy not only minimizes repetitive LLM calls but also improves overall system visibility, providing a cost-effective solution to manage the growing volume of log data. VarParser sets the stage for more insightful system monitoring and failure recovery processes.
