The problem is, what counts as "best"? Letting the optimiser work through all possible plans may well find one with a lower cost, but would it be quicker? It would just be the quickest plan for a mythical machine where the cost of reading from the disk vs. cost of CPU time was what optimiser thought it was and where no relevant data was still in memory. The fastest plan on your cheap PC with 1 hard drive may not be the best for the production server.