Decomposition and Modeling of Microservice Architecture for Distributed Forecasting Systems

Purpose. The aim of the study is to systematize and comparatively analyze methods of decomposition and modeling of microservice architecture for forecasting systems, as well as to substantiate promising directions for the development of this research area with regard to the specific characteristics of analytical workloads. Methodology. The methodological basis is a systemic and comparative analysis of decomposition methods: classification of approaches into static, dynamic, hybrid, and artifact-driven; comparison of graph theory methods, Domain-Driven Design (DDD), and artificial intelligence (NLP, graph neural networks) according to criteria of automation level, domain semantics coverage, and suitability for ML components; formalization of workload heterogeneity. Scientific novelty. Based on a generalization of MLOps practices, the key types of heterogeneity in the analytical workloads are formalized as criteria for decomposition effectiveness: compute asymmetry between ML model training and inference, a dual-loop data lifecycle structure (stateful/stateless), and the isolation of technological dependencies of components. A hybrid three-stage approach is proposed, combining DDD to fix logical domain boundaries, dataflow-oriented optimization to isolate data loops, and automated graph-based and AI methods for the verification of microservice boundaries. Conclusions. It is established that decomposition methods form four classes, among which hybrid approaches demonstrate the best quality yet remain the least common. It is proven that methods with the highest degree of automation (graph-based, AI) are insufficiently effective in terms of compute asymmetry and stateful/stateless separation. Solving the problem of the absence of decomposition methods that account for workload specifics lies in developing hybrid approaches with dynamic verification based on spatiotemporal graph neural networks.