DeQL Studio: Declarative Decision-Making over Relational Data

Abstract

Databases separate what from how: users declare what data they want, and a query optimizer decides how to compute it efficiently. Decision-making has no such separation. Computing optimal actions means solving a mathematical optimization problem, and the user must choose how. That choice is the formulation , and alternative formulations reach the same optimum but can differ in solving time by orders of magnitude. A transportation problem, for instance, can be cast as a general mixed-integer linear program (MILP) or, far faster, as min-cost network flow (MCF). We demonstrate DeQL (Decision Query Language), a declarative SQL extension where users state, over relational data, what to decide, what constraints to respect, and what to optimize; the DeQL Planner chooses the formulation. In DeQL Studio, our interactive interface, attendees write DeQL queries, inspect the formulation the Planner selects, and run conventional scripts for comparison: on a GPU allocation problem with 2.7 million possible assignments, the conventional MILP takes Gurobi 26 s to solve to optimality, while the Planner detects the bipartite supply-and-demand structure directly in the query and selects MCF, and the engine reaches the same optimum in 0.88 s, 30× faster. Attendees edit queries and data, watching the engine adapt: removing one constraint changes the detected structure and the Planner switches formulation, while editing data warm-starts the solver from the previous solution.