On Sat, Mar 30, 2024 at 12:34 PM Tomas Vondra
<tomas.vondra@enterprisedb.com> wrote:
> Hmmm. I admit I didn't think about the "always prefetch" flag too much,
> but I did imagine it'd only affect some places (e.g. BHS, but not for
> sequential scans). If it could be done by lowering the combine limit,
> that could work too - in fact, I was wondering if we should have combine
> limit as a tablespace parameter too.
Good idea! Will add. Planning to commit the basic patch very soon,
I'm just thinking about what to do with Heikki's recent optimisation
feedback (ie can it be done as follow-up, he thinks so, I'm thinking
about that today as time is running short).
> But I think adding such knobs should be only the last resort - I myself
> don't know how to set these parameters, how could we expect users to
> pick good values? Better to have something that "just works".
Agreed.
> I admit I never 100% understood when exactly the kernel RA kicks in, but
> I always thought it's enough for the patterns to be only "close enough"
> to sequential. Isn't the problem that this only skips fadvise for 100%
> sequential patterns, but keeps prefetching for cases the RA would deal
> on it's own? So maybe we should either relax the conditions when to skip
> fadvise, or combine even pages that are not perfectly sequential (I'm
> not sure if that's possible only for fadvise), though.
Yes that might be worth considering, if we know/guess what the OS RA
window size is for a tablespace. I will post a patch for that for
consideration/testing as a potential follow-up as it's super easy,
just for experimentation. I just fear that it's getting into the
realms of "hard to explain/understand" but on the other hand I guess
we already have the mechanism and have to explain it.