Efficient dynamic programming algorithms for ordering expensive joins and selections

Abstract

The generally accepted optimization heuristics of pushing selections down does not yield optimal plans in the presence of expensive predicates. Therefore, several researchers have proposed algorithms to compute optimal processing trees for queries with expensive predicates. All these approaches are incorrect—with one exception [3]. Our contribution is as follows. We present a formally derived and correct dynamic programming algorithm to compute optimal bushy processing trees for queries with expensive predicates. This algorithm is then enhanced to be able to (1) handle several join algorithms including sort merge with a correct handling of interesting sort orders, to (2) perform predicate splitting, to (3) exploit structural information about the query graph to cut down the search space. Further, we present efficient implementations of the algorithms. More specifically we introduce unique solutions for efficiently computing the cost of the intermediate plans and for saving memory space by utilizing bitvector contraction. Our implementations impose no restrictions on the type of query graphs, the shape of processing trees or the class of cost functions. We establish the correctness of our algorithms and derive tight asymptotic bounds on the worst case time and space complexities. We also report on a series of benchmarks showing that queries of sizes which are likely to occur in practice can be optimized over the unconstrained search space in less than a second.

Research supported by the German Research Association (DFG) under contract Mo 507/6-1.