Umeå University's logo

This thesis investigates the cost efficiency of vertical and horizontal scaling approaches in financial software architectures by implementing and evaluating equivalent monolithic and microservice versions of a high-throughput brokerage system. The research examines what bottlenecks limit effective scaling in each architecture and identifies when a monolithic architecture with vertical scaling should be chosen over a microservice architecture with horizontal scaling based on performance-to-cost considerations.

The study implements a financial system handling both IO-intensive order processing and CPU-intensive price update workflows commonly found in brokerage systems. Both architectures are deployed in a Kubernetes environment with equivalent resources. Both incorporating realistic security features like mutual TLS encryption, and are subjected to identical workloads.

Results reveal that neither architecture universally outperforms the other across all scenarios. The monolithic architecture demonstrates significantly higher initial order throughput but struggles to effectively scale both workflows simultaneously, with performance gains heavily dependent on database resources rather than system resources. The microservice architecture, while starting with much lower throughput, scales both workflows more effectively and continuously, but never catches up with the monolithic order processing capability. Service mesh technology significantly impacts microservice performance by adding substantial communication latency and resource consumption.

The findings suggest that monoliths are better suited for simple systems with single dataflows, particularly those leveraging complex database operations, while microservices offer better resource control for multi-dataflow systems despite communication overhead. This research contributes valuable insights for architects and developers designing scalable cloud-native systems, highlighting that architectural decisions should consider specific workflow characteristics rather than following universal best practices.