On Fri, Dec 13, 2002 at 05:00:32PM +0000, Hannu Krosing wrote:
> On Fri, 2002-12-13 at 14:56, george young wrote:
> > > Of course you could just determine all possibly useful indexes
> > > and generate then anyhow an then drop them if they were not used
> > > ;)
> > Why not! At least for selects, this seems like the ideal. For
> > insert and update, you have to deal with updating the superfluous
> > indexes -- does the planner include index updating in its work
> > estimates?
Well, i had a few reasons i didn't want to *actually* create the
1- Disk space. If it's evaluating all indices, including multi-column
indices, that ends up being a significant space drain.
2- Time. Creating indices can take a while for big tables (again,
moreso for multi-column indices).
3- Usability on running systems. If i can eliminate actual index
creation, it won't tie up disk access on systems that are already
dealing with high load.
> At least I think we don't optimize the plan for different index
> access patterns for updating indexes.
I don't think that's the case either, which makes it more difficult to
estimate negative cost of index creation. Not sure how i'll deal with
that except by (for now) ignoring it.
In response to
pgsql-performance by date
|Next:||From: Jean-Luc Lachance||Date: 2002-12-13 16:42:25|
|Subject: Re: [PERFORM] CLUSTER command|
|Previous:||From: johnnnnnn||Date: 2002-12-13 15:20:54|
|Subject: Re: automated index suggestor -- request for comment|