On Thu, Mar 12, 2026 at 9:13 PM Peter Geoghegan <pg@bowt.ie> wrote:
> I'm trying to be practical here. I want to design a solution
> compatible with the changes that we're making to the table AM
> interface. That fixes Mark's complaint. Ideally, this should avoid
> adding much new code to hot code paths. Any design that meets those
> goals is acceptable to me.
The latest version of the index prefetching patch set has a new patch for this:
https://postgr.es/m/CAH2-WzkZTkDuyVFszLwPJesF9pS5E8m0UA+344bx-B-zfA5kaw@mail.gmail.com
My original proposal involved wholly replacing VISITED_PAGES_LIMIT; my
new approach complements it by specifically targeting its one major
weakness. A new INDEX_PAGES_LIMIT mechanism will only tally index leaf
page reads that return zero matching items to the table AM, and will
only give up when that count reaches 3. VISITED_PAGES_LIMIT itself
works as before.
Importantly, this allows VISITED_PAGES_LIMIT to work exactly the same
way on standbys as it does today. During hot standby, LP_DEAD bits
cannot be set or used to avoid heap fetches. As long as each leaf page
returns at least one match (which could still be a TID pointing to a
dead heap tuple), nothing changes -- INDEX_PAGES_LIMIT does nothing
new, we rely on VISITED_PAGES_LIMIT in the traditional way.
--
Peter Geoghegan
