Clustering Models for Analyzing the Database Update Process

This work presents an empirical characterization of Oracle Datapatch execution during live patching by analyzing performance telemetry generated by the Oracle RDBMS. The main problem addressed is the lack of studies, datasets or methodologies describing how Datapatch behaves under real runtime conditions. The general objective of this work is to model this behavior using unsupervised learning techniques in order to identify recurring execution patterns. To achieve this objective, a data acquisition strategy based solely on native and license free instrumentation was developed using Statspack. A structured feature selection process was then applied using two complementary approaches: a statistical method based on a supervised model to identify variables correlated with execution duration, and a semantic method based on Oracle documentation to construct interpretable performance indicators. Both variable sets were used to train and evaluate clustering models capable of grouping similar execution behaviors. The results show that the proposed methodology can distinguish between stable executions, executions with internal pressure spikes handled efficiently and executions characterized by dominant I/O demand. These clusters provide insight into runtime behavior beyond elapsed time measurement and demonstrate that Datapatch performance varies across identifiable operational states. The main contribution of this work is the development of a reproducible methodology for understanding Datapatch execution behavior through data driven analysis. Finally, the conclusions and potential research extensions are presented.