On Wed, May 25, 2011 at 5:54 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Wed, May 25, 2011 at 1:04 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> Because the problem is not specific to TOAST tables. As things
>>> currently stand, we will accept the word of an ANALYZE as gospel even if
>>> it scanned 1% of the table, and completely ignore the results from a
>>> VACUUM even if it scanned 99% of the table. This is not sane.
>
>> I agree that if VACUUM scanned 99% of the table, it's probably fine to
>> use its numbers. It's also fine to use the numbers from ANALYZE,
>> because those pages are chosen randomly. What bothers me is the idea
>> of using a small *non-random* sample, and I'm not sure that
>> incorporating possibly-bogus results slowly is any better than
>> incorporating them quickly.
>
> The above is simply fuzzy thinking. The fact that ANALYZE looked at a
> random subset of pages is *no guarantee whatsoever* that its results are
> highly accurate. They might be more trustworthy than VACUUM's nonrandom
> sample of a similar number of pages, but it doesn't hold even a little
> bit of water to claim that we should believe ANALYZE completely and
> VACUUM not at all even when the latter has looked at a significantly
> larger sample of pages.
I think you're arguing against a straw-man.
> In any case, your line of thought doesn't help us for fixing the problem
> with toast tables, because we aren't going to start doing ANALYZEs on
> toast tables.
Can we simply use a constant for tuple density on TOAST tables?
> The bottom line here is that making use of stats we have is a lot better
> than not making use of them, even if they aren't entirely trustworthy.
http://dilbert.com/strips/comic/2008-05-07/
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
