Characterization of KPI Outliers from Logs Using Data Mining

Ubisoft records the interaction between its customers and its servers in large execution logs (also called traces). Any failure of the system is thus recorded therein. However, the considerable size of these logs considerably hinders their effective use by analysts and developers. We propose an automated method to detect failing executions, and furthermore to characterize the features that are common to clusters of failing instances. The approach will be based an machine learning algorithms, and will produce clusters of failing traces with common features. Since isolating of the features common to a runtime failure is an important part of the overall effort of fixing the issue, the research being proposed here will allow developers to extract actionable information from the traces.

Faculty Supervisor:

Raphaël Khoury

Student:

Souhail Ben Ali

Partner:

Ubisoft

Discipline:

Computer science

Sector:

Information and cultural industries

University:

Université du Québec à Chicoutimi