Re: index prefetching

Hi,

On Mon, 21 Jul 2025 at 03:59, Thomas Munro <thomas.munro@gmail.com> wrote:
>
> On Sun, Jul 20, 2025 at 1:07 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> > On Sat, Jul 19, 2025 at 11:23 PM Tomas Vondra <tomas@vondra.me> wrote:
> > > The thing that however concerns me is that what I observed was not the
> > > distance getting reset to 1, and then ramping up. Which should happen
> > > pretty quickly, thanks to the doubling. In my experiments it *never*
> > > ramped up again, it stayed at 1. I still don't quite understand why.
> >
> > Huh.  Will look into that on Monday.
>
> I suspect that it might be working as designed, but suffering from a
> bit of a weakness in the distance control algorithm, which I described
> in another thread[1].  In short, the simple minded algorithm that
> doubles on miss and subtracts one on hit can get stuck alternating
> between 1 and 2 if you hit certain patterns.  Bilal pinged me off-list
> to say that he'd repro'd something like your test case and that's what
> seemed to be happening, anyway?  I will dig out my experimental
> patches that tried different adjustments to escape from that state....

I used Tomas Vondra's test [1]. I tracked how many times
StartReadBuffersImpl() functions return true (IO is needed) and false
(IO is not needed, cache hit). It returns true ~%6 times on both
simple and complex patches (~116000 times true, ~1900000 times false
on both patches).

A complex patch ramps up to ~250 distance at the start of the stream
and %6 is enough to stay at distance. Actually, it is enough to ramp
up more but it seems the max distance is about ~270 so it stays there.
On the other hand, a simple patch doesn't ramp up at the start of the
stream and %6 is not enough to ramp up. It is always like distance is
1 and IO needed, so multiplying the distance by 2 -> distance = 2 but
then the next block is cached, so decreasing the distance by 1 and
distance is 1 again.

[1] https://www.postgresql.org/message-id/aa46af80-5219-47e6-a7d0-7628106965a6%40vondra.me

--
Regards,
Nazir Bilal Yavuz
Microsoft