johnnnnnn wrote:

On Wed, Jun 25, 2003 at 10:30:31AM -0500, Andrew Dunstan wrote:

DB2 looks good. I have horrid, horrid memories of wrestling with the
Oracle "extent" madness.

I do think that it's worth providing additional access points to tablespaces, though. That is, it would make sense to me to allow "CREATE INDEX indexname IN spacename", instead of attaching an indexspace to a table.

This is especially true with postgresql, since i've seen more than one
proposal for multi-table indices. If we're spacing indices based on
the table, it's unclear where a given multi-table index should go.

It would also allow for other flexibilities, like putting join indices
(on foreign keys) in one tablespace, with indices for aggregation or
sorting in another tablespace.

I wonder why an index spanning multiple tables should be stored in a different location than the tables itself. If we're talking about derived tables, all data/index must be available at the same time to be meaningful, so why not restrict them to the same tablespace? This sounds like more flexibility than really useful to me.

The philosophy of pgsql is to let the os and the io system distribute the load over disks and other resources, not to do it in the backend. That's why we need much less organizational effort than other systems that try to implement everything themselves, on raw devices etc.


---------------------------(end of broadcast)--------------------------- TIP 8: explain analyze is your friend

Reply via email to