On 6/18/21 9:54 PM, John Naylor wrote:
>
> On Fri, Jun 18, 2021 at 3:43 PM Tomas Vondra
> <tomas.vondra@enterprisedb.com <mailto:tomas.vondra@enterprisedb.com>>
> wrote:
>
> > Sorry, I'm not sure what you mean by "we set the number of MCVs to the
> > number of histograms" :-(
> >
> > When you say "MCV limit" you mean that we limit the number of items to
> > statistics target, right? I agree plan time is one concern - but it's
> > also about analyze, as we need larger sample to build a larger MCV or
> > histogram (as the paper you referenced shows).
>
> Ah, I didn't realize the theoretical limit applied to the MCVs too, but
> that makes sense since they're basically singleton histogram buckets.
>
Something like that, yes. Looking at MCV items as singleton histogram
buckets is interesting, although I'm not sure that was the reasoning
when calculating the MCV size. AFAIK it was kinda the other way around,
i.e. the sample size is derived from the histogram paper, and when
building the MCV we ask what's sufficiently different from the average
frequency, based on the sample size etc.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company