CAN HTAP ELIMINATE ETL? AN EMPIRICAL ANALYSIS OF OLAP, OLTP, AND HTAP SYSTEMS

The traditional architectural divide between Online Transaction Processing (OLTP) and Online Analytical Processing (OLAP) systems has historically led organizations to rely heavily on Extract-Transform-Load (ETL) pipelines to bridge transactional databases with analytical data warehouses. While effective for batch analytics, this decoupled architecture introduces latency, reduces data freshness, and increases operational complexity, especially for modern applications that demand real-time insights.

Hybrid Transactional/Analytical Processing (HTAP) database systems aim to unify OLTP and OLAP capabilities within a single platform, potentially eliminating the need for dedicated ETL pipelines. This study presents a comprehensive comparative analysis of OLTP (MySQL with row-based storage), OLAP (Amazon Redshift with column-based storage), and HTAP (TiDB with hybrid storage) systems. Industry-standard benchmarks, including TPC-H for mixed workloads and the Star Schema Benchmark (SSB) for analytical workloads, are used to evaluate performance across various schema designs (star vs. flat), scale factors (SF=5, SF=10, SF=20), and cluster sizes. Key evaluation metrics include query execution time, mean latency, throughput, data freshness rate, and scalability.

The findings reveal that HTAP systems such as TiDB not only match or exceed the analytical performance of dedicated OLAP systems but also provide real-time data availability by eliminating traditional ETL delays. Moreover, TiDB's hybrid storage model and distributed architecture enable efficient handling of complex workloads, offering dynamic query optimization and strong consistency guarantees.

Based on these insights, this study argues that HTAP systems represent a viable, unified alternative to ETL-dependent architectures. It offers recommendations for their adoption in enterprise environments seeking to simplify infrastructure, reduce latency, and improve responsiveness to real-time data.