RE: sequential read on full-table scan?

  • From: Martic Zoran <zoran_martic@xxxxxxxxx>
  • To: Christian.Antognini@xxxxxxxxxxxx
  • Date: Sun, 22 May 2005 02:32:46 -0700 (PDT)


Thanks for your good answer.

It is always the balance between good and bad things.

The main idea of my questions is to drive more brains
thinking about something possible interesting (at
least to me, I hope to others too).

It looks to me that Oracle is waiting for the response
from the customer base before trying to figure out
what is the best.
I am not sure that they have up to now all
intelligence of what will be the best way to progress
more with system statistics.
My main concern is that these statistics should be at
least per file basis, but of course if object spans
multiple files or object is partitioned it brings you
more dimensions.
But as Oracle oversimplified and used avearages for
everything why it should not use averages per file
basis too, at least for objects belongs to the
specific file.

Oh sorry, maybe tablespace level may be even better.

Per object may be very big price to be paid in the
level of code and complexity. But not sure that it is
crazy hard per tablespace, especially because Oracle
has file stats already that can be used to be averaged
over tablespaces on which objects are based.

Let's wait what is going to happen in 10g R2 if any :)

At the end the system statistics that are bringing the
intelligence about I/O subsystem performances are the
key piece in getting proper optimizer COST and good
execution plan. Disk I/O are the slowest part of the
modern systems so making sure that knowledge about it
may be more important then something else.


--- Christian Antognini
<Christian.Antognini@xxxxxxxxxxxx> wrote:
> Hi Zoran
> >The main question is how good is that prediction
> based
> >on the global system statistics only.
> >Not sure how these statistics if collected per file
> or
> >object level will be more useful???
> If you want to have really good statistics you
> should add either a time =
> dimension and some kind of forecasting (regression?)
> or real-time =
> statistics.
> Of course both have major problems... e.g.: What
> about segments that =
> span more files? What about partitioned objects?
> When are the cached =
> execution plans invalidated?=20
> IMHO it's better to KISS.
> Just my 2 cents!
> Chris
> --

Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 

Other related posts: