Context: Technical debt (TD) has emerged as an important research topic in software engineering, reflecting the long-term negative consequences of suboptimal design decisions. A significant effort has been made towards TD identification, classification, prioritization, and management. Nevertheless, empirical evidence on how TD evolves in industrial, large-scale software systems and how it relates to architectural complexity and delivery performance remains scarce.

Objective: The objective of this thesis is to develop an empirical understanding of the evolution of TD in large-scale software systems and its relationship with architectural complexity indicators and software delivery performance. Understanding TD evolution and these relationships is crucial for making informed decisions regarding maintenance, refactoring, and resource allocation.

Methodology: This thesis includes empirical case studies employing a mixed methods research approach, combining longitudinal case studies with qualitative insights from practitioners. We collected quantitative data on TD, complexity indicators, lead time, and size by mining software repositories using tools such as SonarQube and Jira. Qualitative data were collected through structured discussions, semi-structured interviews, and focus groups. To analyze quantitative data, we used robust statistical methods, while qualitative data were analyzed thematically and mapped to the quantitative findings.

Results: The findings indicate that TD exhibits distinct temporal patterns across the architectural styles studied. Furthermore, complexity indicators generally increase over time and co-evolve with TD. The thesis identifies four interconnected architectural challenge areas that contextualize the observed trends in architectural complexity. Finally, extending the investigation to software delivery performance, no consistent association is observed between TD and lead time across the studied components.

Conclusions: TD evolution is a complex, context-dependent phenomenon influenced by architectural, organizational, and human factors, as well as technical aspects such as code quality. The findings emphasize the importance of continuous monitoring, context-aware interpretation, and the integration of both technical and socio-technical perspectives to effectively manage TD in practice.