On Wed, Jan 18, 2006 at 08:13:59PM -0500, Tom Lane wrote:
> Simon Riggs <simon@2ndquadrant.com> writes:
> > We only need to index the row with the lowest value on any page so the main
> > index would get 100 times smaller. The main part of the index would not
> > need to be written to except when a block overflows.
>
> BTW, the above is equivalent to saying that the leaf-level index pages
> aren't there: the downlink pointers on the level-1 index pages are
> pointers to heap pages, instead, and you're right that they effectively
> only index the lowest value per page (actually IIRC the highest value
> per page, but same difference).
Would this open the door for allowing tables to be maintained in CLUSTER
order (at least at the block level if not within the blocks)? Though I
have no idea how you'd handle page splits without a lot of pain, but
perhaps it would be possible to strive for a certain tuple ordering that
would allow for a periodic re-cluster that doesn't have to move a lot of
data. One thought is to strive for the same amount of free space on each
page, so if you're touching a tuple on a page that has less than desired
free space you move it's new version to either the next or previous
page.
--
Jim C. Nasby, Sr. Engineering Consultant jnasby@pervasive.com
Pervasive Software http://pervasive.com work: 512-231-6117
vcard: http://jim.nasby.net/pervasive.vcf cell: 512-569-9461