Random Sampling from Databases

Random Sampling from Databases by Frank Olken Doctor of Philosophy in Computer Science University of California at Berkeley Professor Michael Stonebraker, Chair In this thesis I describe e cient methods of answering random sampling queries of relational databases, i.e., retrieving random samples of the results of relational queries. I begin with a discussion of the motivation for including sampling operators in the database management system (DBMS). Uses include auditing, estimation (e.g., approximate answers to aggregate queries), and query optimization. The second chapter contains a review of the basic le sampling methods used in the thesis: acceptance/rejection sampling, reservoir sampling, and partial sum (ranked) tree sampling. I describe their usage for sampling from variably blocked les, and sampling from results as they are generated. Related literature on sampling from databases is reviewed. In Chapter Three I show how acceptance/rejection sampling ofB trees can be employed to obtain simple random samples of B tree les without auxiliary data structures. Iterative and batch algorithms are described and evaluated. The fourth chapter covers sampling from hash les: open addressing hash les, separately chained over ow hash les, linear hash les, and extendible hash les. I describe both iterative and batch algorithms, and characterize their performance. I describe and analyze algorithms for sampling from relational operators in Chapter Five: selection, intersection, union, projection, set di erence, and join. Methods of sampling from complex relational expressions, including select-project-join queries, are also described. In Chapter Six I describe the maintenance of materialized sample views. Here I combine sampling techniques with methods of maintaining conventional materialized views. I consider views de ned by simple queries consisting of single relational operators. The penultimate chapter covers sampling from spatial databases. I develop algorithms for obtaining uniformly distributed samples of points which satisfy a spatial predicate represented as a union of polygons in the database. Sampling algorithms from both quadtrees and R-trees are described, including spatial reservoir sampling algorithms. I conclude with a summary of the thesis and an agenda for future work.