On Fri, Sep 11, 2020 at 7:57 AM Matthias van de Meent <boekewurm+postg...@gmail.com> wrote: > I've not yet been involved in postgresql's development process, so here's a > first. Please find attached a patch for improving the BT binary search > speeds, with accompanying performance data. > > v1-0001 adapts _bt_compare and _bt_binsrch* to make use of static properties > of L&Y-style nbtree indexes to speed up finding an initial position in the > nbtee.
Are you familiar with this thread? https://www.postgresql.org/message-id/cah2-wzn_nayk4pr0hrwo0stwhmxjp5qyu+x8vppt030xpqr...@mail.gmail.com I wrote a patch that implemented the same idea, which is sometimes called dynamic prefix truncation. I found some very subtle problems with it, though. Concurrent page deletions could occur, which could cause a scan that skips a prefix to miss that the page it landed on doesn't have the same common prefix anymore. > there could be a large throughput increase for short strings if the collation > lookup from lc_collate_is_c() in varstr_cmp could be reused in the subsequent > call to pg_newlocale_from_collation() I noticed that myself recently. -- Peter Geoghegan
