Sorguların çalışma süresinin tahmini için yükle ilişkili öznitelik mühendisliği

Prediction of query execution time is one of the most challenging issues for relational databases and is useful for database administration, resource management, system monitoring and query scheduling. Most of the query optimizers use cost-based models for query execution time prediction but the problem is more complex because the heterogeneity of the database system's hardware platforms and operating systems makes more difficult to measure CPU and I/O costs. The relational database vendors try to implement autonomous databases which automates management and performance thus intelligent query execution time prediction is a key issue. Previous work mostly used synthetical data so that reproducing machine learning experiments are almost impossible for various domains. In this thesis, we use real-world data of a payment service provider with different workloads and we propose new sets of features based on aggregating the database queries and compared them with traditional query plan features. We collected data from a common machine data tool so that reproducing ma-chine learning experiments and building models are easy for various domains.