On Fri, Jul 28, 2023 at 5:03 PM Melanie Plageman
<melanieplageman@gmail.com> wrote:
> When you were working on this, what were the downsides of only having the
> criteria that 1) page would be all_visible/all_frozen and 2) we did prune
> something (i.e. no consideration of FPIs)?
Conditioning freezing on "all_visible && all_frozen" seems likely to
always be a good idea when it comes to any sort of speculative trigger
criteria. Most of the wins in this area will come from avoiding future
FPIs, and you can't hope to do that unless you freeze everything on
the page. The cost of freezing when "all_visible && all_frozen" does
not hold may be quite low, but the benefits will also be very low.
Also, the way that recovery conflicts are generated for freeze records
somewhat depends on this right now -- though you could probably fix
that if you had to.
Note that we *don't* really limit ourselves to cases where an FPI was
generated from pruning, exactly -- even on 16. With page-level
checksums we can also generate an FPI just to set a hint bit. That
triggers freezing too, even on insert-only tables (assuming checksums
are enabled). I expect that that will be an important condition for
triggering freezing in practice.
Your question about 2 seems equivalent to "why not just always
freeze?". I don't think that that's a bad question -- quite the
opposite. Even trying to give an answer to this question would amount
to getting involved in new work on VACUUM, though. So I don't think I
can help you here.
--
Peter Geoghegan
