Re: MVCC catalog access

Robert Haas escribió:

> All right, so here's a patch that does something along those lines.
> We have to always take a new snapshot when somebody scans a catalog
> that has no syscache, because there won't be any invalidation messages
> to work off of in that case. The only catalog in that category that's
> accessed during backend startup (which is mostly what your awful test
> case is banging on) is pg_db_role_setting. We could add a syscache
> for that catalog or somehow force invalidation messages to be sent
> despite the lack of a syscache, but what I chose to do instead is
> refactor things slightly so that we use the same snapshot for all four
> scans of pg_db_role_setting, instead of taking a new one each time. I
> think that's unimpeachable on correctness grounds; it's no different
> than if we'd finished all four scans in the time it took us to finish
> the first one, and then gotten put to sleep by the OS scheduler for as
> long as it took us to scan the other three. Point being that there's
> no interlock there.

That seems perfectly acceptable to me, yeah.

> The difference is 3-4%, which is quite a lot less than what you
> measured before, although on different hardware, so results may vary.

3-4% on that synthetic benchmark sounds pretty acceptable to me, as
well.

--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services