> > Whereas if you set toast_tuples_per_page to 8k then the only option
> > for Postgres will be to put each datum in its own page and
> waste 1-3k
> > on every page.
>
> No, because actually the code is designed to make the toast
> chunk size just enough less than 8K that the tuples fit.
He mentioned toasted values of 6-7k each. If all values are of that size
(like payment slip tiffs) there is nothing we would do with the
remaining 1-2k on each page. But that disadvantage disappears as soon as
you have just a little more variability in the length. Still, it might
be enough to opt for some freespace reuse smarts if we can think of a
cheap heuristic. But the cost to fetch such a "distributed" tuple would
be so huge I doubt there is anything to win but disk space.
Andreas