LogFlow: Simplified Log Analysis for Large Scale Systems
Marc Platini, Thomas Ropars, Benoît Pelletier, Noël De Palma · 2020
Distributed infrastructures generate huge amount of logs that can provide useful information about the state of system, but that can be challenging to analyze. The paper presents LogFlow, a tool to help human operators in the analysis of logs by automatically constructing graphs of correlations between log entries. The core of LogFlow is an interpretable predictive model based on a Recurrent Neural Network augmented with a state-of-the-art attention layer from which correlations between log entries are deduced. To be able to deal with huge amount of data, LogFlow also relies on a new log parser algorithm that can be orders of magnitude faster than best existing log parsers. Experiments run with several system logs generated by Supercomputers and Cloud systems show that LogFlow is able to achieve more than 96% of accuracy in most cases.