RE: Performance (was: The New Slashdot Setup (includes MySql server))

> -----Original Message-----
> From: Tom Lane [mailto:tgl(at)sss(dot)pgh(dot)pa(dot)us]
>
> "Hiroshi Inoue" <Inoue(at)tpf(dot)co(dot)jp> writes:
> >> Unfortunately there is no index on pg_index's indrelid column in 7.0,
> >> so this is not fixable without an initdb. TODO item for 7.1, I guess.
>
> I've also been thinking about ways to implement the relcache-based
> caching of index information that I mentioned before. That doesn't
> address the scanning problem in general but it should improve
> performance for this part of the planner quite a bit.

Sounds reasonble to me.

> The trick is to
> ensure that other backends update their cached info whenever an index
> is added or deleted. I thought of one way to do that: force an update
> of the owning relation's pg_class tuple during CREATE or DROP INDEX,
> even when we don't have any actual change to make in its contents ---
> that'd force a relcache invalidate cycle at other backends. (Maybe
> we don't even need to change the pg_class tuple, but just send out a
> shared-cache-invalidate message as if we had.)
>

Seems CREATE INDEX already sends shared_cache_invalidate
message. DROP INDEX doesn't ?? I'm not sure.

> > I know another case. pg_attrdef has no index on (adrelid,attnum)
> > though it has an index on (adrelid).
>
> Doesn't look to me like we need an index on (adrelid,attnum), at
> least not in any paths that are common enough to justify maintaining
> another index. The (adrelid) index supports loading attrdef data
> into the relcache, which is the only path I'm particularly concerned
> about performance of...
>

It seems to me that an index on (adrelid,adnum) should
exist instead of the current index. It identifies pg_attrdef.
I say *Oops* about it in my trial implementation of ALTER
TABLE DROP COLUMN.

Regards.

Hiroshi Inoue
Inoue(at)tpf(dot)co(dot)jp