Re: eliminate xl_heap_visible to reduce WAL (and eventually set VM on-access)

Thanks for all the reviews. I'm working on responding to your previous
mails with a new version.

On Wed, Aug 27, 2025 at 8:55 AM Kirill Reshke <reshkekirill(at)gmail(dot)com> wrote:
>
> v6-0015:
> I chose to verify whether this single modification would be beneficial
> on the HEAD.
>
> Benchmark I did:
>
> ```
>
> \timing
> CREATE TABLE zz(i int);
> alter table zz set (autovacuum_enabled = false);
> TRUNCATE zz;
> copy zz from program 'yes 2 | head -n 180000000';
> copy zz from program 'yes 2 | head -n 180000000';
>
> delete from zz where (REPLACE(REPLACE(ctid::text, '(', '{'), ')',
> '}')::int[])[2] = 7 ;
>
> VACUUM FREEZE zz;
> ```
>
> And I checked perf top footprint for last statement (vacuum). My
> detailed results are attached. It is a HEAD vs HEAD+v6-0015 benchmark.
>
> TLDR: function inlining is indeed beneficial, TransactionIdPrecedes
> function disappears from perf top footprint, though query runtime is
> not changed much. So, while not resulting in query speedup, this can
> save CPU.
> Maybe we can derive an artificial benchmark, which will show query
> speed up, but for now I dont have one.

I'm not surprised that vacuum freeze does not show a speed up from the
function inlining.

This patch was key for avoiding a regression in the most contrived
worst case scenario example of setting the VM on-access. That is, if
you are pruning only a single tuple on the page as part of a SELECT
query that returns no tuples (think SELECT * FROM foo OFFSET N where N
is greater than the number of rows in the table), and I add
determining if the page is all visible, then the overhead of these
extra function calls in heap_prune_record_unchanged_lp_normal() is
noticeable.

We might be able to come up with a similar example in vacuum without
freeze since it will try to determine if the page is all-visible. Your
example is still running on my machine, though, so I haven't verified
this yet :)

- Melanie