Change-Impact-Guided Failure Localization in Continuous Deployment Pipelines

Abstract

Failures introduced during continuous deployment are often difficult to localize because a single release may contain changes spanning application code, configuration files, dependencies, and infrastructure definitions. This paper proposes a change-impact-guided method for locating deployment-related failures by connecting release changes with runtime symptoms. The method first constructs dependency relationships among modified components and affected services, and then combines change proximity, failure timing, and runtime anomaly signals to rank potential failure sources. To reduce false associations, historical deployments are used to estimate whether particular components frequently exhibit similar runtime variations after successful releases. The approach is evaluated using deployment histories containing configuration errors, dependency incompatibilities, resource regressions, and application-level defects. Compared with ranking methods based only on temporal correlation or service dependencies, the proposed method identifies the responsible component within the top-ranked candidates more consistently and reduces the number of components engineers need to inspect. The results suggest that deployment context can provide useful evidence for practical failure localization without requiring fully automated root-cause reasoning.