Journey to find abnormal activities from logging data with periodic patterns
Background
Monitoring and configuring alerts for a new Single Sign-On service was made easy, but detecting abnormal activities using traditional metrics remains challenging. To address this issue, we utilized Facebook Prophet, a time series forecasting library, for a proof-of-concept project. Our project focused on forecasting session information for the new open-source-based Single Sign-On Service.
By using the past 28 days of data, we were able to forecast 3 hours of data. Previously, only linear alarm settings were possible. However, with the forecasting data, we can now set fitting alarm settings for every time.
