Re: New GUC autovacuum_max_threshold ?

Hi frederic.yhuel

> Thank you. FWIW, I would prefer a sub-linear growth, so maybe something
> like this

> vacthresh = Min(vac_base_thresh + vac_scale_factor * reltuples,
> vac_base_thresh + vac_scale_factor * pow(reltuples, 0.7) * 100);

> This would give :

> * 386M (instead of 5.1 billion currently) for a 25.6 billion tuples
table ;
> * 77M for a 2.56 billion tuples table (Robert's example) ;
> * 15M (instead of 51M currently) for a 256M tuples table ;
> * 3M (instead of 5M currently) for a 25.6M tuples table.
> The other advantage is that you don't need another GUC.
Argee ,We just need to change the calculation formula，But I prefer this
formula because it calculates a smoother value.

vacthresh = (float4) fmin(vac_base_thresh + vac_scale_factor *
reltuples,vac_base_thresh
+ vac_scale_factor * log2(reltuples) * 10000);
or
vacthresh = (float4) fmin(vac_base_thresh + (vac_scale_factor * reltuples)
, sqrt(1000.0 * reltuples));

Frédéric Yhuel <frederic(dot)yhuel(at)dalibo(dot)com> 于2024年8月12日周一 21:41写道：

>
>
> On 8/7/24 23:39, Nathan Bossart wrote:
> > I've attached a new patch to show roughly what I think this new GUC
> should
> > look like. I'm hoping this sparks more discussion, if nothing else.
> >
>
> Thank you. FWIW, I would prefer a sub-linear growth, so maybe something
> like this:
>
> vacthresh = Min(vac_base_thresh + vac_scale_factor * reltuples,
> vac_base_thresh + vac_scale_factor * pow(reltuples, 0.7) * 100);
>
> This would give :
>
> * 386M (instead of 5.1 billion currently) for a 25.6 billion tuples table ;
> * 77M for a 2.56 billion tuples table (Robert's example) ;
> * 15M (instead of 51M currently) for a 256M tuples table ;
> * 3M (instead of 5M currently) for a 25.6M tuples table.
>
> The other advantage is that you don't need another GUC.
>
> > On Tue, Jun 18, 2024 at 12:36:42PM +0200, Frédéric Yhuel wrote:
> >> By the way, I wonder if there were any off-list discussions after
> Robert's
> >> conference at PGConf.dev (and I'm waiting for the video of the conf).
> >
> > I don't recall any discussions about this idea, but Robert did briefly
> > mention it in his talk [0].
> >
> > [0] https://www.youtube.com/watch?v=RfTD-Twpvac
> >
>
> Very interesting, thanks!
>
>
>