Log clustering based problem identification for online service systems

Logs play an important role in the maintenance of large-scale online service systems. When an online service fails, engineers need to examine recorded logs to gain insights into the failure and identify the potential problems. Traditionally, engineers perform simple keyword search (such as "error" and "exception") of logs that may be associated with the failures. Such an approach is often time consuming and error prone. Through our collaboration with Microsoft service product teams, we propose LogCluster, an approach that clusters the logs to ease log-based problem identification. LogCluster also utilizes a knowledge base to check if the log sequences occurred before. Engineers only need to examine a small number of previously unseen, representative log sequences extracted from the clusters to identify a problem, thus significantly reducing the number of logs that should be examined, meanwhile improving the identification accuracy. Through experiments on two Hadoop-based applications and two large-scale Microsoft online service systems, we show that our approach is effective and outperforms the state-of-the-art work proposed by Shang et al. in ICSE 2013. We have successfully applied LogCluster to the maintenance of many actual Microsoft online service systems. In this paper, we also share our success stories and lessons learned.