Selected problems in cardinality estimation

Müller, Magnus

PDF
Dissertation_Magnus_Mueller.pdf - Veröffentlichte Version
Download (4MB)

URN:	urn:nbn:de:bsz:180-madoc-636836
Dokumenttyp:	Dissertation
Erscheinungsjahr:	2022
Ort der Veröffentlichung:	Mannheim
Hochschule:	Universität Mannheim
Gutachter:	Moerkotte, Guido
Datum der mündl. Prüfung:	21 Dezember 2022
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Practical Computer Science III (Moerkotte 1996-)
Lizenz:	Creative Commons Namensnennung 4.0 International (CC BY 4.0)
Fachgebiet:	004 Informatik
Freie Schlagwörter (Englisch):	cardinality estimation , query optimization
Abstract:	Cardinality estimation remains a critical task in query processing. Query optimizers rely on the accuracy of cardinality estimates when generating execution plans, and, in approximate query answering, estimated cardinalities affect the quality of query results. In this thesis, we present multiple new cardinality estimation techniques. The techniques differ vastly by the query under consideration. For single relation queries, we use the principle of maximum entropy to combine information extracted from samples and histograms. For join size estimation, we rely on a model that requires one to find estimates for the intersection size of join attributes. For queries with multiple joins, sketches serve as compact representations of join results that are combined via a data structure that approximates the joint frequency distribution of join attributes. In addition, we present a technique to transform selection predicates into a representation that allows estimators based on machine learning to effectively learn query result cardinalities. For each cardinality estimator presented in this thesis, we precisely define its problem scope, the construction process, and how to obtain estimates. Then, we compare to state-of-the-art cardinality estimators and run a thorough evaluation with queries over multiple data sets. Based on our observations, we analyze the strengths and limitations of each of our cardinality estimators and identify its preferred use case.