Processamento de dados e modelagem de séries temporais para detecção de anomalias em arquiteturas de microsserviço

Abstract

Microservices architecture has gained widespread adoption due to its scalability and flexibility, but it also poses significant challenges for monitoring and fault detection. Manual analysis and simple statistical methods often result in false positives or negatives, hindering the accurate identification of anomalies. This work proposes a comprehensive methodology for anomaly detection in microservices environments through time series metric processing and the application of deep learning. As there was no access to real data, a plausible synthetic database was developed containing HTTP latency, Kafka lag, and CPU usage metrics, reproducing real problems such as different collection frequencies, temporal gaps, outliers, and decentralized distribution of metrics. Next, a preprocessing pipeline was developed involving temporal standardization, format transformation, missing data imputation, and information enrichment, among others. Simulated anomalies—representing common crashes in production—were then inserted into the dataset. Finally, an initial model based on 1D CNN was trained and evaluated for the task of binary classification between normal and anomalous behavior. The results obtained showed good precision, recall, and F1-score values, indicating the potential of the approach, although limited by the use of artificial data. The study demonstrates the feasibility of applying deep learning techniques for anomaly detection in distributed systems and lays the groundwork for future work with real data and more refined models