Re: reducing random_page_cost from 4 to 2 to force index scan

On Sun, May 15, 2011 at 2:08 PM, Josh Berkus <josh@agliodbs.com> wrote:
>> All true.  I suspect that in practice the different between random and
>> sequential memory page costs is small enough to be ignorable, although
>> of course I might be wrong.
>
> This hasn't been my experience, although I have not carefully measured
> it.  In fact, there's good reason to suppose that, if you were selecting
> 50% of more of a table, sequential access would still be faster even for
> an entirely in-memory table.
>
> As a parallel to our development, Redis used to store all data as linked
> lists, making every object lookup effectively a random lookup.  They
> found that even with a database which is pinned in memory, creating a
> data page structure (they call it "ziplists") and supporting sequential
> scans was up to 10X faster for large lists.
>
> So I would assume that there is still a coefficient difference between
> seeks and scans in memory until proven otherwise.

Well, anything's possible.  But I wonder whether the effects you are
describing might result from a reduction in the *number* of pages
accessed rather than a change in the access pattern.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company