On Fri, Apr 19, 2013 at 7:43 PM, Jeff Janes <jeff.janes@gmail.com> wrote:
> On Fri, Apr 19, 2013 at 2:24 PM, Claudio Freire <klaussfreire@gmail.com>
> wrote:
>>
>>
>> Especially if there's some locality of occurrence, since analyze
>> samples pages, not rows.
>
>
> But it doesn't take all rows in each sampled page. It generally takes about
> one row per page, specifically to avoid the problem you indicate. Maybe it
> is possible to trick it into taking too many (for example, if most pages are
> completely empty), but I haven't experienced that as being a problem.
Still, I remember a discussion where it was clear there was a bias
towards sampling rows from the same page (or nearby), resulting in
this particular problem.