An Efficient Implementation of Game Provenance in DBMS

Explaining why a certain answer is in the result ofa query or why it is missing from the result is important for manyapplications including auditing, debugging data, and answeringhypothetical questions about data. Both types of questions, i.e.,why provenance and why-not (missing answer) provenance havebeen studied extensively. Provenance games, a game-theoreticapproach to provenance, can provide a unified view on why andwhy-not provenance. The power of provenance games, however,comes at the cost of complexity: a direct approach for computinggame provenance instantiates a large game graph with all possibletuples that can be formed from values of the active domainusing the rules of a query evaluation game. In this work, wepresent a new relational database approach for computing amuch smaller subgraph of the provenance game that is relevantto the user’s provenance question. In addition to avoiding theevaluation of a non-stratified win-move Datalog program, we alsoexclude parts of the game graph early on if we can determinethat they will not be relevant to explain the user question. Wepresent an implementation in the GProM provenance databasemiddleware. Our experimental evaluation demonstrates that theapproach scales to large instances and significantly outperformsthe direct method.