CoDS: A Representative Sampling Method for Relational Databases

Abstract

Abstract. Database sampling has become a popular approach to handle large amounts of data in a wide range of application areas such as data mining or approximate query evaluation. Using database samples is a potential solution when using the entire database is not cost-effective, and a balance between the accuracy of the results and the computational cost of the process applied on the large data set is preferred. Existing sampling approaches are either limited to specific application areas, to single table databases, or to random sampling. In this paper, we propose CoDS: a novel sampling approach targeting relational databases that ensures that the sample database follows the same distribution for specific fields as the original database. In particular it aims to maintain the distribution between tables. We evaluate the performance of our algorithm by measuring the representativeness of the sample with respect to the original database. We compare our approach with two existing solutions, and we show that our method performs faster and produces better results in terms of representativeness.

ReX: Extrapolating Relational Data in a Representative Way

“…The sizes of the tables range from 77 (District) to 1,056,320 tuples (Trans). The Financial database schema is depicted in [4]. The starting table identified by ReX is the District table.…”

“…Both ReX and UpSizeR aim to scale the distributions of the relationships between tables by s (i.e., through primary and foreign keys). In [4] we proposed a sampling method that aimed to scale the same distributions by a sampling factor. We use the average representativeness error metric defined in [4], replacing the sampling rate with the scaling rate.…”

“…In [4] we proposed a sampling method that aimed to scale the same distributions by a sampling factor. We use the average representativeness error metric defined in [4], replacing the sampling rate with the scaling rate. Moreover, we use the global size error metric defined in [4] to evaluate the size of X related to O.…”

“…Thus, the method must decide for which partial number of tuples of t j it should create new tuples. As this represents a di↵erent problem by itself [4,8], in this paper we consider only natural scaling rates. Moreover, the scenario of naturally scaling databases is commonly applicable to enterprises where it is rarely needed to extrapolate to a fraction rather than a natural number.…”

“…In this paper, we propose an automated representative extrapolation technique, ReX, that addresses the scaling problem above. Similarly to [4] and [19], we define a representative database as a database where the distributions of the relationships between the tables are preserved from the original database. As foreign keys are enforced links between tables, they represent invaluable inputs to depict the relationships between data in a relational database.…”