Scalable Software Engineering Architecture for AI-Enabled ETL Pipelines Using Event-Driven Microservices
As data ecosystems become more diverse and time-critical, traditional monolithic ETL pipelines face challenges to meet the demands of modern data engineering workloads in terms of scalability, adaptability, and operational resilience. In this paper, we introduce an event-driven microservices approach to orchestrate and deploy AI-based ETL (ETL = Extraction, Transformation, and Loading) pipelines in a Kubernetes-managed environment that includes the following components: asynchronous orchestration using Apache Kafka, hybrid anomaly detection, adaptive schema inference, and predictive load balancing. The proposed architecture breaks the ETL processing into loosely coupled services, which can be deployed and scaled independently, and incorporates data quality intelligence into the transformation layer. Two explicit baselines are used for experimental evaluation: (1) a traditional monolithic batch-processing ETL pipeline with sequential execution of the stages performed without horizontal scaling, and (2) an event-driven microservices pipeline, which uses Apache Kafka orchestration but no AI-enabled optimization modules. The framework is able to achieve 39.7% improvement in throughput (4,820 vs. 3,450 records/sec) and 96.3% anomaly detection accuracy, while reducing average E2E latency by 14.6% (245 vs. 287 ms). The framework's throughput is 288.7% higher than the baseline of ETL monolithic, and the average latency drops 72.5% to 245 ms compared to 892 ms for the baseline. An architectural assessment also shows that the system is more modular, loosely coupled, more fault isolated, and more flexible to deploy; all of which are important software quality attributes. The results indicate the proposed architecture as a potentially reusable reference to create scalable and intelligent ETL systems in enterprise data processing environments and emphasize the necessity of generalization in multi-domain validation.