On Wed, Sep 9, 2020 at 3:47 AM David Rowley <dgrowle...@gmail.com> wrote: > On Tue, 8 Sep 2020 at 12:08, Thomas Munro <thomas.mu...@gmail.com> wrote: > > One thought is that if we're going to copy everything out and back in > > again, we might want to consider doing it in a > > memory-prefetcher-friendly order. Would it be a good idea to > > rearrange the tuples to match line pointer order, so that the copying > > work and also later sequential scans are in a forward direction? > > That's an interesting idea but wouldn't that require both the copy to > the separate buffer *and* a qsort? That's the worst of both > implementations. We'd need some other data structure too in order to > get the index of the sorted array by reverse lineitem point, which > might require an additional array and an additional sort.
Well I may not have had enough coffee yet but I thought you'd just have to spin though the item IDs twice. Once to compute sum(lp_len) so you can compute the new pd_upper, and the second time to copy the tuples from their random locations on the temporary page to new sequential locations, so that afterwards item ID order matches offset order.
