On Tue, Jul 27, 2010 at 9:56 PM, Tom Lane <tgl@xxxxxxxxxxxxx> wrote: > Robert Haas <robertmhaas@xxxxxxxxx> writes: >> On Tue, Jul 27, 2010 at 4:40 PM, Tom Lane <tgl@xxxxxxxxxxxxx> wrote: >>> "Flushing them all" is not zero-cost; it's not too hard to believe that >>> it could actually be slower than forking a clean new backend. > >> I'm not so sure I believe that. > > I'm not asserting it's true, just suggesting it's entirely possible. > Other than the fork() cost itself and whatever authentication activity > there might be, practically all the startup cost of a new backend can be > seen as cache-populating activities. You'd have to redo all of that, > *plus* pay the costs of getting rid of the previous cache entries. > Maybe the latter costs less than a fork(), or maybe not. fork() is > pretty cheap on modern Unixen. > >>> What's much worse, it's not zero-bug. > >> I think this is probably the biggest issue with the whole idea, and I >> agree there would be some pain involved. > > Yeah, if it weren't for that I'd say "sure let's try it". But I'm > afraid we'd be introducing significant headaches in return for a gain > that's quite speculative. I agree that the gain is minimal of itself; after all, how often do you need to switch databases, and what's the big deal if the postmaster has to fork a new backend? Where I see it as a potentially big win is when it comes to things like parallel query. I can't help thinking that's going to be a lot less efficient if you're forever forking new backends. Perhaps the point here is that you'd actually sort of like to NOT flush all those caches unless it turns out that you're switching databases - many installations probably operate with essentially one big database, so chances are good that even if you distributed connections / parallel queries to backends round-robin, you'd potentially save quite a bit of overhead. Of course, for the guy who has TWO big databases, you might hope to be a little smarter, but that's another problem altogether. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise Postgres Company -- Sent via pgsql-performance mailing list (pgsql-performance@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
