EXPLAINABLE CNN-BASED CLASSIFICATION OF HETEROGENEOUS DATABASE WORKLOADS USING RUNTIME PERFORMANCE FEATURES

The current database management systems handle heterogeneous workloads, such as transactional, analytical, reporting, and mixed workloads that have varied patterns of execution and resource demands. Proper classification of workloads can be used to monitor performance of the database and optimize it based on workload. In the paper, a explainable Convolutional Neural Network (CNN)-based model is proposed to classify heterogeneous database workloads based on runtime performance characteristics, such as query execution time, CPU use, memory use, disk input/output activity, query complexity and parallelist. The features gathered are pre-processed and they are fed into a CNN model to determine four workload categories. SHapley Additive exPlanations (SHAP) is used to enhance interpretability, by determining the features that contributed to individual predictions and overall classification behavior. This framework is assessed in the form of accuracy, precision, recall, F1-score, classification latency, and computational overhead. It is compared to conventional classifiers, such as Decision Tree, Random Forest, Support Vector Machine, and Multilayer Perceptron. TPC-H and TPC-DS benchmarks are assumed to be used with analytical workloads, which are complemented by other workloads that model the other categories. The proposed framework will help to offer the correct and understandable workload classification and establish the basis of future adaptive query processing and optimization of the database.