Inconsistent time for a.get_nns_by_item(i,n) #665
Description
When running a.get_nns_by_item(i,n) on a fixed AnnoyIndex already built, the time to find the item is sometimes about 2 times longer than usual for a fixed i . This is shown in the linked png of a histogram of the time it takes to do the computation. I would expect 1 peak, but not the secondary smaller one at about twice the time
