On Wed, 2008-05-28 at 16:55 -0400, Gregory Stark wrote:
> "Simon Riggs" <simon@2ndquadrant.com> writes:
>
> > So the idea is to have one pass per VACUUM, but make that one pass do
> > the first pass of *this* VACUUM and the second pass of the *last*
> > VACUUM.
>
> I think that's exactly the same as the original suggestion of having HOT
> pruning do the second pass of the last vacuum. The trick is to know whether
> the last vacuum committed or not. If it didn't commit then it's not safe to
> remove those line pointers yet.
Perhaps, though I'm not suggesting storing extra xids on-block.
I think if we have to wait for a VACUUM to run before marking the line
pointers then we may as well wait for two. Having something wait for a
VACUUM and then removed it by HOT afterwards gives you the worst of both
worlds: long wait for a VACUUM then more overhead and extra code during
HOT pruning.
-- Simon Riggs www.2ndQuadrant.comPostgreSQL Training, Services and Support