Horton
Menu

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

Melanie Plageman <melanieplageman@gmail.com>

View thread
From: Melanie Plageman <melanieplageman@gmail.com>
To: Andres Freund <andres@anarazel.de>
Cc: Andrey Borodin <x4mmm@yandex-team.ru>, Kirill Reshke <reshkekirill@gmail.com>, Chao Li <li.evan.chao@gmail.com>, Xuneng Zhou <xunengzhou@gmail.com>, Robert Haas <robertmhaas@gmail.com>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>, Heikki Linnakangas <hlinnaka@iki.fi>
Date: 2026-03-03T00:04:24Z
Lists: pgsql-hackers 
On Fri, Feb 20, 2026 at 4:34 PM Andres Freund <andres@anarazel.de> wrote:
>
> On 2026-01-28 18:16:10 -0500, Melanie Plageman wrote:
> > Subject: [PATCH v34 13/14] Allow on-access pruning to set pages all-visible
> >
> > Many queries do not modify the underlying relation. For such queries, if
> > on-access pruning occurs during the scan, we can check whether the page
> > has become all-visible and update the visibility map accordingly.
> > Previously, only vacuum and COPY FREEZE marked pages as all-visible or
> > all-frozen.
> >
> > This commit implements on-access VM setting for sequential scans as well
> > as for the underlying heap relation in index scans and bitmap heap
> > scans.
>
> For evaluating this, did you build anything that evaluates the frequency of
> this succeeding, causing unnecessary un-all-visibling etc during benchmarks?

I didn't develop a specific micro-benchmark for this, but I did run
some generic pgbenches (which does a single tuple update on accounts
followed by a select) because I thought there would be a good amount
of un-all-visibling there. I didn't gather stats to confirm though and
who knows with a random data distribution (IIRC it was a relatively
small working set, but still). I can develop something more targeted,
though.

> > @@ -631,7 +632,9 @@ heap_prepare_pagescan(TableScanDesc sscan)
> >       /*
> >        * Prune and repair fragmentation for the whole page, if possible.
> >        */
> > -     heap_page_prune_opt(scan->rs_base.rs_rd, buffer);
> > +     if (sscan->rs_flags & SO_HINT_REL_READ_ONLY)
> > +             vmbuffer = &scan->rs_vmbuffer;
> > +     heap_page_prune_opt(scan->rs_base.rs_rd, buffer, vmbuffer);
>
> I don't love that the signalling to heap_page_prune_opt() about this is by
> passing vmbuffer or NULL.

v35 is more explicit and heap_page_prune_opt() has a rel_read_only flag.

> We clearly don't want to actually freeze rows if we're doing an update and
> might just update the rows again. But it's less clear to me that, if we are
> pruning dead row versions *and* the page is already all-visible after that
> (say because only HOT versions were removed), we shouldn't mark the page as
> such?

If we're doing an update and the new tuple fits on the same page, then
the page will not be all-visible by the time the update is over,
right? And if the new tuple doesn't fit on the same page as the old
tuple, then while it would be nice to mark the old page as
all-visible, don't we on-access prune the page before actually
updating the tuple? Like we are scanning in the old page to update it
and on-access prune then to make space for it and then we make the
page modification.

> > @@ -306,6 +312,13 @@ heap_page_prune_opt(Relation relation, Buffer buffer)
> >                               .cutoffs = NULL,
> >                       };
> >
> > +                     if (vmbuffer)
> > +                     {
> > +                             visibilitymap_pin(relation, BufferGetBlockNumber(buffer), vmbuffer);
> > +                             params.options |= HEAP_PAGE_PRUNE_UPDATE_VM;
> > +                             params.vmbuffer = *vmbuffer;
>
> Why do we pin the buffer at this time, rather than deferring that until we
> actually need it?  I guess we just always will access it, but that doesn't
> seem like it's inherent (c.f. my earlier points about a faster exit when
> looking at an already all-frozen page or such).

We would need to pin the VM to see if it is all-frozen to exit early.
For the on-access case, since we won't freeze, we could rely on
PD_ALL_VISIBLE to exit early, but that means we wouldn't be able to
identify and fix PD_ALL_VISIBLE/VM-all-visible mismatches.

> It's not clear to me why we are pinning the page in lazy_scan_heap(), before
> it's clear that we need it, either.  But there the cost is often very low,
> because we have a lot of sequential accesses.  But here we might be called
> from an index scan, with very little locality of access.

Now that, as of v35, we check for VM corruption unconditionally at the
start of heap_page_prune_and_freeze() and check the VM to potentially
exit early, there's no benefit in deferring pinning the VM in either
vacuum or on-access.

- Melanie

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Stabilize btree_gist test against on-access VM setting

  2. Stabilize plancache test against on-access VM setting

  3. Make local buffers pin limit more conservative

  4. Update FSM during prune/freeze replay even if freespace is zero

  5. Pass down information on table modification to scan nodes

  6. Thread flags through begin-scan APIs

  7. Remove PlannedStmt->resultRelations in favor of resultRelationRelids

  8. Make it cheap to check if a relation is modified by a query

  9. Keep newest live XID up-to-date even if page not all-visible

  10. Add pruning fast path for all-visible and all-frozen pages

  11. Detect and fix visibility map corruption in more cases

  12. Save vmbuffer in heap-specific scan descriptors for on-access pruning

  13. Avoid BufferGetPage() calls in heap_update()

  14. Initialize missing fields in CreateExecutorState()

  15. Prefix PruneState->all_{visible,frozen} with set_

  16. Add PageGetPruneXid() helper

  17. Move commonly used context into PruneState and simplify helpers

  18. Remove table_scan_analyze_next_tuple unneeded parameter OldestXmin

  19. Simplify visibility check in heap_page_would_be_all_visible()

  20. Rename LVRelState VM-related logging counters

  21. Fix flakiness in the pg_visibility VM-only vacuum test by using a temporary table.

  22. Refactor lazy_scan_prune() VM clear logic into helper

  23. Eliminate use of cached VM value in lazy_scan_prune()

  24. Combine visibilitymap_set() cases in lazy_scan_prune()

  25. Fix const qualification in prune_freeze_setup()

  26. Simplify vacuum visibility assertion

  27. Split heap_page_prune_and_freeze() into helpers

  28. Assert that cutoffs are provided if freezing will be attempted

  29. Split PruneFreezeParams initializers to one field per line

  30. Refactor heap_page_prune_and_freeze() parameters into a struct

  31. Make heap_page_is_all_visible independent of LVRelState

  32. Inline TransactionIdFollows/Precedes[OrEquals]()

  33. Add helper for freeze determination to heap_page_prune_and_freeze

  34. Bump XLOG_PAGE_MAGIC after xl_heap_prune change

  35. Correct prune WAL record opcode name in comment

  36. Add error codes when vacuum discovers VM corruption

  37. Remove unused xl_heap_prune member, reason

  38. Remove unneeded VM pin from VM replay

  39. Add assert and log message to visibilitymap_set

  40. Stamp HEAD as 19devel.

  41. Add error codes to some corruption log messages