Anomaly Detection for ETL Execution Graphs Using Temporal Dependency Modeling

Abstract

Extract, Transform, and Load (ETL) pipelines support data movement, cleansing, validation, and integration across modern enterprise systems. However, abnormal execution behavior in ETL workflows can cause delayed processing, incomplete data loading, dependency failures, and unreliable analytical outputs. This article presents an anomaly detection approach for ETL execution graphs through temporal dependency modeling. The proposed method represents ETL tasks, execution paths, task dependencies, runtime duration, failure sequence, and scheduling behavior as time-aware graph structures. By modeling both structural relationships and temporal execution patterns, the framework identifies unusual delays, broken dependencies, repeated task failures, and abnormal workflow transitions. The approach improves visibility into hidden execution risks that are difficult to detect using rule-based monitoring alone. The study shows that temporal dependency modeling can support early failure detection, better root-cause interpretation, and improved reliability of largescale ETL systems. This work contributes to intelligent data pipeline monitoring and resilient enterprise data engineering.