Multi-tuple deletion propagation

This paper studies the computational complexity of the classic problem of deletion propagation in a relational database, where tuples are deleted from the base relations in order to realize a desired deletion of tuples from the view. Such an operation may result in a (sometimes unavoidable) side effect: deletion of additional tuples from the view, besides the intentionally deleted ones. The goal is to minimize the side effect. The complexity of this problem has been well studied in the case where only a single tuple is deleted from the view. However, only little is known within the more realistic scenario of multi-tuple deletion, which is the topic of this paper. The class of conjunctive queries (CQs) is among the most well studied in the literature, and we focus here on views defined by CQs that are self-join free (sjf-CQs). Our main result is a trichotomy in complexity, classifying all sjf-CQs into three categories: those for which the problem is in polynomial time, those for which the problem is NP-hard but polynomial-time approximable (by a constant-factor), and those for which even an approximation (by any factor) is NP-hard to obtain. A corollary of this trichotomy is a dichotomy in the complexity of deciding whether a side-effect-free solution exists, in the multi-tuple case. We further extend the full classification to accommodate the presence of a constant upper bound on the number of view tuples to delete, and the presence of functional dependencies. Finally, we establish (positive and negative) complexity results on approximability for the dual problem of maximizing the number of view tuples surviving (rather than minimizing the side effect incurred in) the deletion propagation.