Mark Kirkwood <markir@paradise.net.nz> writes:
> Tom Lane wrote:
>> With this model, the disk cost to fetch a single
>> index entry will be estimated as random_page_cost (default 4.0) rather
>> than the current fixed 2.0. This shouldn't hurt things too much for
>> simple indexscans --- especially since anyone running with a reduced
>> random_page_cost won't see as much change. And it will increase the costs
>> for bitmap scans that cross many index pages, which is what we need in
>> light of Philippe's numbers.
> This sounds good to me, although the 2.0 -> 4.0 cost jump may cause
> (more) cases of people seeing their index scans in pre-8.2 versions
> becoming some other type of access in 8.2. I guess a comment about
> testing existing applications could be placed in the 8.2 release notes?
Yeah, that comes with the territory. One point to note is that with
this model, setting random_page_cost below 2.0 will actually make small
indexscans look *cheaper* than they do now. So it'll certainly be
possible to make the thing jump in that direction if you need to.
regards, tom lane