The impact of concept drift and data leakage on log level prediction models

Abstract

Developers insert logging statements to collect information about the execution of their systems. Along with a logging framework (e.g., Log4j), practitioners can decide which log statement to print or suppress by tagging each log line with a log level. Since picking the right log level for a new logging statement is not straightforward, machine learning models for log level prediction (LLP) were proposed by prior studies. While these models show good performances, they are still subject to the context in which they are applied, specifically to the way practitioners decide on log levels in different phases of the development history of their projects (e.g., debugging vs. testing). For example, Openstack developers interchangeably increased/decreased the verbosity of their logs across the history of the project in response to code changes (e.g., before vs after fixing a new bug). Thus, the manifestation of these changing log verbosity choices across time can lead to concept drift and data leakage issues, which we wish to quantify in this paper on LLP models. In this paper, we empirically quantify the impact of data leakage and concept drift on the performance and interpretability of LLP models in three large open-source systems. Additionally, we compare the performance and interpretability of several time-aware approaches to tackle time-related issues. We observe that both shallow and deep-learning-based models suffer from both time-related issues. We also observe that training a model on just a window of the historical data (i.e., contextual model) outperforms models that are trained on the whole historical data (i.e., all-knowing model) in the case of our shallow LLP model. Finally, we observe that contextual models exhibit a different (even contradictory) model interpretability, with a (very) weak correlation between the ranking of important features of the pairs of contextual models we compared. Our findings suggest that data leakage and concept drift should be taken into consideration for LLP models. We also invite practitioners to include the size of the historical window as an additional hyperparameter to tune a suitable contextual model instead of leveraging all-knowing models.