Diagnosing failures in large-scale distributed systems often relies solely on operational logs, as failures are difficult to reproduce and stack traces are frequently missing or unreliable. Existing bug localization techniques assume the availability ...
Diagnosing failures in large-scale distributed systems often relies solely on operational logs, as failures are difficult to reproduce and stack traces are frequently missing or unreliable. Existing bug localization techniques assume the availability of structured bug reports, making them unsuitable for real-world production settings where only raw logs remain. This thesis presents a log-based bug localization framework that integrates crash-marker log slicing, semantic code retrieval using vector embeddings, and structured reasoning with large language models (LLMs). We construct a dataset of 274 log–bug pairs from 13 open-source distributed systems and evaluate the proposed method using standard retrieval metrics. The framework achieves a file-level Recall@10 of 51.5%, demonstrating that meaningful bug localization is possible even without bug reports or stack traces. Direct LLM prompting with raw logs fails due to token limits and irrelevant context, confirming the necessity of combining log slicing and retrieval. These results provide empirical evidence that operational logs alone can support automated bug localization and lay the groundwork for log-centric debugging and reliability engineering in modern distributed systems.