Complex SQL queries are widely used today, but it is rather difficult to check if a complex query has been written correctly. Formal verification based on comparing a specification with an implementation is not applicable, since SQL queries are essentially a specification without any implementation. Queries are usually checked by running them on sample datasets and checking that the correct result is returned; there is no guarantee that all possible errors are detected. In this paper, we address the problem of test data generation for checking correctness of SQL queries, based on the query mutation approach for modeling errors. Our presentation focuses in particular on a class of join/outer-join mutations, comparison operator mutations, and aggregation operation mutations, which are a common cause of error. To minimize human effort in testing, our techniques generate a test suite containing small and intuitive test datasets. The number of datasets generated, is linear in the size of the query, although the number of mutations in the class we consider is exponential. Under certain assumptions on constraints and query constructs, the test suite we generate is complete for a subclass of mutations that we define, i.e., it kills all non-equivalent mutations in this subclass.

Checking if an SQL query has been written correctly is not an easy task. Formal verification is not applicable, since it is based on comparing a specification with an implementation, whereas SQL queries are essentially a specification without any implementation. Thus, the standard approach for testing queries is to manually check query results on test datasets. Intuitively, a mutant is a query variant that could have been the correct query if the query was in error; a mutant is killed by a dataset if the original query and the mutant return different results on the dataset. In this paper, we address the problem of generation of test data for an SQL query, to kill mutants. Our work focuses in particular on a class of join/outer-join mutants, which are a common cause of error. To minimize human effort in testing, our techniques generate a test suite containing small and intuitive test datasets, combining them into a single dataset where possible. In the absence of foreign-key constraints, and under certain assumptions, the test suite is complete, i.e. it kills all nonequivalent mutations, in the class of join-type mutations that we consider. We also consider some common types of where-clause predicate mutants. Our techniques have been implemented in a prototype data generation tool.