On Mon, Oct 6, 2025 at 11:12:00AM +0200, Tomas Vondra wrote:
> On 10/6/25 11:02, Michael Banck wrote:
> > Hi,
> >
> > On Mon, Oct 06, 2025 at 02:59:16AM +0200, Tomas Vondra wrote:
> >> I started looking at how we calculated the 4.0 default back in 2000.
> >> Unfortunately, there's a lot of info, as Tom pointed out in 2024 [2].
> >> But he outlined how the experiment worked:
> >>
> >> - generate large table (much bigger than RAM)
> >> - measure runtime of seq scan
> >> - measure runtime of full-table index scan
> >> - calculate how much more expensive a random page access is
> >
> > Ok, but I also read somewhere (I think it might have been Bruce in a
> > recent (last few years) discussion of random_page_cost) that on top of
> > that, we assumed 90% (or was it 95%?) of the queries were cached in
> > shared_buffers (probably preferably the indexes), so that while random
> > access is massively slower than sequential access (surely not 4x by
> > 2000) is offset by that. I only quickly read your mail, but I didn't see
> > any discussion of caching on first glance, or do you think it does not
> > matter much?
> >
>
> I think you're referring to this:
>
> https://www.postgresql.org/message-id/1156772.1730397196%40sss.pgh.pa.us
>
> As Tom points out, that's not really how we calculated the 4.0 default.
> We should probably remove that from the docs.
Uh, that might not be how we tested to find the default value, but I
thought it was the logic of why we _thought_ it was such a low value
compared to the speed difference of magnetic random vs sequential I/O.
--
Bruce Momjian <bruce@momjian.us> https://momjian.us
EDB https://enterprisedb.com
Do not let urgent matters crowd out time for investment in the future.
