Predicting ERP Latency with Machine Learning : A Validated Telemetry Pipeline for Diagnostic and Predictive Data Analysis

Sammanfattning

ERP systems are sensitive to latency increases, because they cause significant operational consequences. Although these systems generate large volumes of telemetry data, transforming it into a reliable, analysis-ready dataset to be used for optimization, remains challenging. This study designs and evaluates a telemetry pipeline for an ERP system, asking whether it produces high quality data and whether that data supports ML-based latency analysis. A three-layer monitoring system was built using Windows Exporter, SQL Anywhere Exporter, and Locust, with Prometheus for collection and DuckDB for offline analysis. Data was collected across eight synthetic load scenarios at two-second resolution and validated automatically for structural integrity, temporal continuity, and value plausibility. Three ML models were then evaluated against a persistence baseline at a 120-second forecasting horizon. The pipeline achieved a 100% validation pass rate. CPU utilisation showed the strongest correlation with p95 latency (r = 0.557), followed by memory and database concurrency signals. Tree-based ML models reduced MAE by 35–38% over the baseline, with the most predictive features drawn from engineered application-layer signals. The results showed that a purpose-built telemetry pipeline can produce analysis-ready ERP performance data supporting both diagnostic and ML-based latency prediction. The system sustained zero request failures despite substantial latency degradation, highlighting the value of cross-layer monitoring even in nominally stable systems.