I've no idea if this is related but I recall finding that on windows 10 copying the rowids from a very large index into a vector was much much slower if I accessed the index in descending order. I discovered that during the descending order copy windows available memory gradually decreased as if windows was caching the entire index. The memory usage and time difference between the asc/desc copies was huge. Keith Medcalf was able to reproduce it and contacted ms but nothing was ever done about it. The thread is on the old Nabble forum here http://sqlite.1065341.n5.nabble.com/timer-td102512.html If you think it might be related and look it up I'd start at the post Keith got involved as many of my earlier posts / experiments were flawed.