On Tue, 2008-10-28 at 14:28 +0200, Heikki Linnakangas wrote:
> Hannu Krosing wrote:
> > On Tue, 2008-10-28 at 10:10 +0000, Simon Riggs wrote:
> >> On Tue, 2008-10-28 at 11:45 +0200, Hannu Krosing wrote:
> >>> On Tue, 2008-10-28 at 08:49 +0000, Simon Riggs wrote:
> >>>> Looking at a VACUUM's WAL records makes me think twice about the way we
> >>>> issue a VACUUM.
> >>>> 1. First we scan the heap, issuing a HEAP2 clean record for every block
> >>>> that needs cleaning.
> >>> IIRC the first heap pass just collects info and does nothing else.
> >>> Is this just an empty/do-nothing WAL record ?
> >> 8.3 changed that; it used to work that way. I guess I never looked at
> >> the amount of WAL being generated.
> > I can't see how it is safe to do anything more than just lookups on
> > first pass.
> What's done in the first pass is the same HOT pruning that is done
> opportunistically on other page accesses as well. IIRC it's required for
> correctness, though I can't remember what exactly the issue was.
Are you sure it is a correctness thing ? Maybe HOT pruning just happened
to be in a path used by vacuum to read pages.
> I don't think the extra WAL volume is a problem;
Probably not ( unless you need to ship your WAL records via a very
expensive network connection ).
If it is a simple performance problem, then it can probably be fixed by
just running VACUUM slower.
> VACUUM doesn't generate
> much WAL, anyway. As for the extra data page writes it causes; yeah,
> that might cause some I/O that could be avoided, but remember that the
> first pass often dirties buffers anyway to set hint bits.
Still, can't we special-case HOT pruning and hint-bit change WAL-logging
for first the pass of vacuum ? They both seem redundant in case of
In response to
pgsql-hackers by date
|Next:||From: Tom Lane||Date: 2008-10-28 12:55:07|
|Subject: Re: Window Functions: v07 APIs and buffering strateties |
|Previous:||From: Heikki Linnakangas||Date: 2008-10-28 12:44:57|
|Subject: Re: Proposal of PITR performance improvement for 8.4.|