We discuss a class of tightly feasible MILP for which branchand-bound is ineffective. We consider its hardness, evaluate the probability that randomly generated instances are feasible, and introduce a heuristic solution method based on the old idea of reformulating binary variables to continuous while introducing a linear complementarity constraint. We show the extent of the computational advantage, under a time limit, of our heuristic with respect to a state-of-the-art branch-andbound implementation.