On Tue, 2008-10-28 at 10:59 +0000, Gregory Stark wrote:
> Simon Riggs <simon@2ndQuadrant.com> writes:
>
> > I don't see a reason why we would issue 2 WAL records per block for a
> > VACUUM, nor why we would prune and remove in two steps, dirtying the
> > block each time. Seems like we could write approximately half the amount
> > of data that we do.
> >
> > Surely we can come up with a better plan than that one?
>
> This sounds like the same issue Pavan identified and proposed solving by
> rotating the three passes so that we do step 3 at the beginning of the next
> vacuum run, so it can be done in the same pass as step 1.
>
> To do that he proposed we do:
>
> 1. scan heap doing two things: a) remove any marked tuples if they were marked
> by a previous vacuum which committed and b) prune and mark any tuples we
> find are deletable for a future vacuum to remove.
>
> 2. scan indexes and remove the tuples we marked in 1b.
It's fairly hard to remove the second heap pass completely.
I think what I am suggesting is two heap passes, but writing WAL and
dirtying blocks on only one of the passes.
The biggest I/O cost comes from the writes, not the reads, ISTM.
-- Simon Riggs www.2ndQuadrant.comPostgreSQL Training, Services and Support