
104 Automated Physical Database Design and Tuning
picked by the optimizer to evaluate any query in the workload whenever the
original candidate indexes are also present. Additionally, indexes usually in-
teract with each other, and therefore assigning a fixed benefit value to each
index will miss opportunities (see our discussion in the previous section in the
context of greedy(m,B)). Finally, the simple heuristic to solve the knapsack
problem is efficient but can be twice as bad as the optimal solution, and we
might want to trade additional optimization time with the quality of the re-
sulting configuration. We next comment on some approaches to address ...