A composite latency model for evaluating hybrid OLTP/OLAP information systems

Hybrid transactional and analytical processing (HTAP) systems require simultaneous optimisation of storage, batch loading, and analytics latency components – objectives that frequently conflict under varying data scales. Existing approaches evaluate individual latency metrics in isolation, providing limited guidance for practitioners selecting database configurations in legacy SQL Server environments. This study proposes Lₜₒₜₐₗ, a workload–profile–aware composite latency model that aggregates storage latency (Ls), ETL batch latency (Ll), and BI analytics latency (La) via Simple Additive Weighting (SAW) with configurable weight profiles. We define four progressive configurations (C0–C3) representing DBMS–agnostic architectural archetypes, and evaluate them across three data scales (100K, 1M, 10M rows) using the Olist Brazilian e–commerce dataset on SQL Server 2022. Measurements follow Sequential Layer Latency Evaluation (SLLE), a deliberate design choice reflecting legacy maintenance–window deployment patterns. Each configuration–scale–metric combination was measured over 30 isolated runs after warm–up exclusion (1,080 observations total), with Welch's t–test and Cohen's d for statistical validation. Results reveal a scale–dependent crossover: the baseline heap configuration (C0) minimises Lₜₒₜₐₗ at 100K rows across all workload profiles, while the columnstore–enhanced configuration (C2) achieves 22.8% lower Lₜₒₜₐₗ at production scale (10M rows), driven by a 79% reduction in ETL batch latency (8,208 ms to 1,725 ms). Sensitivity analysis across 28 weight combinations confirms this conclusion in 78.6% of the weight space. A baseline comparison shows that no single latency metric correctly identifies the optimal configuration, underscoring the need for composite evaluation. A two–component mixture model attributes the bimodal analytics–latency distribution in columnstore configurations to a deterministic, selectivity–driven mechanism, and an illustrative licensing–cost scenario shows the composite ranking is robust to a realistic Standard–to–Enterprise price differential but reverts below a cost weight of ≈0.14. The framework enables quantitative, workload–aware configuration selection without schema migration or hardware replacement.