On Thu, 14 Mar 2019 at 23:06, David Rowley <david.rowley@2ndquadrant.com> wrote:
> On Thu, 14 Mar 2019 at 21:12, Edmund Horner <ejrh00@gmail.com> wrote:
> > I'm not sure how an unreasonable underestimation would occur here. If
> > you have a table bloated to say 10x its minimal size, the estimator
> > still assumes an even distribution of tuples (I don't think we can do
> > much better than that). So the selectivity of "ctid >= <last ctid
> > that would exist without bloat>" is still going to be 0.9.
>
> Okay, think you're right there. I guess the only risk there is just
> varying tuple density per page, and that seems no greater risk than we
> have with the existing stats.
Yeah that is a risk, and will probably come up in practice. But at
least we're not just picking a hardcoded selectivity any more.
> Just looking again, I think the block of code starting:
>
> + if (density > 0.0)
>
> needs a comment to mention what it's doing. Perhaps:
>
> + /*
> + * Using the average tuples per page, calculate how far into
> + * the page the itemptr is likely to be and adjust block
> + * accordingly.
> + */
> + if (density > 0.0)
>
> Or some better choice of words. With that done, I think 0001 is good to go.
Ok, I'll look at it and hopefully get a new patch up soon.
Edmund
