DynTrackr: A Robust Two-Stage Framework with Attribute Enhancement for KPI Anomaly Detection

Real-time monitoring of Key Performance Indicators (KPIs) is crucial in software testing, enabling the identification and diagnosis of issues during canary deployments or limited-scale testing phases. Efficient anomaly detection in software instrumentation data is crucial for maintaining software quality. However, the volatility of software telemetry data makes achieving stable and accurate anomaly detection in real-world applications highly challenging. To address these challenges, we propose DynTrackr, a two-stage framework that incorporates a time series forecasting structure with a multi-channel sparse correlation module and a multi-period sequential decomposition module. Additionally, it includes a streamlined anomaly detection mechanism with trend decomposition and difference analysis blocks following the forecasting structure. Dynamic real-time anomaly detection is achieved by integrating a pre-trained forecasting model with basic data imputation in the first stage, and applying straightforward yet effective trend and difference feature extraction in the second stage. Results from experiments conducted on three benchmark datasets and one real-world dataset illustrate the superior performance of DynTrackr over state-of-the-art methods in both univariate and multivariate forecasting, as well as anomaly detection tasks.