Integration of machine learning into high-load web application architectures

Abstract

The article analyzes the feasibility of implementing machine learning in high-load web application architectures.  An architectural technique for integrating machine learning (ML) modules into microservice high-load Kubernetes-based web applications is considered. The basis is the "Sidecar + event-driven" pattern, which allows each service to be supplemented with load forecasting on LSTM models and anomaly detection through auto-encoders. The ML Predictor, Decision Manager, and Actuator components are embedded as separate containers that communicate via the Kafka event bus. The methodological basis of the work, which made it possible to broadly consider the features of the machine learning implementation process in high-load web application architectures, was based on the results of other studies. The information presented in the article is of considerable interest to architects of distributed systems and DevOps engineers specializing in building fault-tolerant, scalable high-traffic web platforms that need to integrate machine learning models without performance degradation and while maintaining SLA requirements. In addition, the materials of the article will be useful to researchers in the field of MLOps and data engineers involved in optimizing data pipelines and ensuring low latency of inference under extreme loads, as well as graduate students developing methods for adaptive resource balancing in heterogeneous computing environments.