Tablespace and cpu costs

On Wed, 18 Aug 2004, Gaetano Mendola wrote:

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi all,
now that we have the tablespace support don't you think that
each tablespace needs his own costs instead of a system wide
one ?

I searched in the archives and on the TODO list without success.

Yep. I've looked at this but didn't think it necessary for the first stage
of implementation.

I definately plan to do it for 8.1

Gavin

Gavin Sherry wrote:

On Wed, 18 Aug 2004, Gaetano Mendola wrote:

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi all,
now that we have the tablespace support don't you think that
each tablespace needs his own costs instead of a system wide
one ?

I searched in the archives and on the TODO list without success.

Yep. I've looked at this but didn't think it necessary for the first stage
of implementation.

I definately plan to do it for 8.1

Do you have an Idea on how you want implement this ( usage side ) ?

Regards
Gaetano Mendola

On Wed, 18 Aug 2004, Gaetano Mendola wrote:

Gavin Sherry wrote:

On Wed, 18 Aug 2004, Gaetano Mendola wrote:

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi all,
now that we have the tablespace support don't you think that
each tablespace needs his own costs instead of a system wide
one ?

I searched in the archives and on the TODO list without success.

Yep. I've looked at this but didn't think it necessary for the first stage
of implementation.

I definately plan to do it for 8.1

Do you have an Idea on how you want implement this ( usage side ) ?

Well, as far as I can tell there are two ways we could do it without
reinventing the wheel -- but both are ugly.

The first is to add a parameter to CREATE TABLESPACE and ALTER TABLESPACE
called RANDOMPAGECOST (or something like that) which allows users to set
the random page cost for a given tablespace. I'm undecided as to whether
they should be able to set the sequential scan cost. That seems
reasonable too at first look.

The second option is to have a pg_tablespace.conf file which defines
tablespace specific settings in some way or another.

Either way, the data will be feed into the planner so that the costing
routines will be able to use them.

Anyone else got some thoughts on how we could do this in a cleaner manner?

Gavin

The first is to add a parameter to CREATE TABLESPACE and ALTER TABLESPACE
called RANDOMPAGECOST (or something like that) which allows users to set
the random page cost for a given tablespace. I'm undecided as to whether
they should be able to set the sequential scan cost. That seems
reasonable too at first look.

The second option is to have a pg_tablespace.conf file which defines
tablespace specific settings in some way or another.

Either way, the data will be feed into the planner so that the costing
routines will be able to use them.

Anyone else got some thoughts on how we could do this in a cleaner manner?

Yes, exactly like we have ALTER USER SET blah=foo; and ALTER DATABASE
SET blah = foo;

Chris

Gavin Sherry wrote:

On Wed, 18 Aug 2004, Gaetano Mendola wrote:

Gavin Sherry wrote:

On Wed, 18 Aug 2004, Gaetano Mendola wrote:

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi all,
now that we have the tablespace support don't you think that
each tablespace needs his own costs instead of a system wide
one ?

I searched in the archives and on the TODO list without success.

Yep. I've looked at this but didn't think it necessary for the first stage
of implementation.

I definately plan to do it for 8.1

Do you have an Idea on how you want implement this ( usage side ) ?

Well, as far as I can tell there are two ways we could do it without
reinventing the wheel -- but both are ugly.

The first is to add a parameter to CREATE TABLESPACE and ALTER TABLESPACE
called RANDOMPAGECOST (or something like that) which allows users to set
the random page cost for a given tablespace. I'm undecided as to whether
they should be able to set the sequential scan cost. That seems
reasonable too at first look.

The second option is to have a pg_tablespace.conf file which defines
tablespace specific settings in some way or another.

Either way, the data will be feed into the planner so that the costing
routines will be able to use them.

That is the goal.

Anyone else got some thoughts on how we could do this in a cleaner manner?

I think we need both, I think could be also usefull add a STORE option:

ALTER TABLESPACE RANDOMPAGECOST 0.001 STORE

if the STORE is present then the pg_tablespace.conf is updated

this could be argued with the fact that we do not have SET cpu_costs STORE bu this
doesn't mean that we can not have it in the future :-)

I think is also usefull insert it in the dump "as comment" before the tablespace
creation, this in order to don't loose these setting.

Regards
Gaetano Mendola