> > But you could do the indexes first and remember how far you
> can vacuum
> > the heap later.
>
> But the indexes _can't_ be done first; you _first_ need to
> know which tuples are dead, which requires looking at the
> table itself.
If we already had the "all tuples visible" bitmap I think we could
first scan the bitmap and decide whether we can afford to look
at the visibility info for each entry in the index.
We only collect the ctids before so we don't have the inefficient
lookups,
but if we can avoid the lookup in most cases it would again be
attractive.
Andreas