Without gainsaying anything earlier, I wonder about this: > Alternatively, loading the data in sorted order and then creating the additional index (after the data is loaded) also works and is the only alternative for a WITHOUT ROWID table to avoid B-Tree rebalance operations. I would think that the insert order would need to be close enough to or match a breadth-first traversal of the ultimate (as) balanced (as it's going to get) B-tree. Otherwise, lots of rebalancing will be needed because, ultimately the ordering will match a depth-first traversal but insertion in that order would necessarily create imbalances. (I'll be happy to be educated on this if I've got it wrong.)