Horton

Re: Emit fewer vacuum records by reaping removable tuples during pruning

Peter Geoghegan <pg@bowt.ie>

View thread View on postgr.es
From: Peter Geoghegan <pg@bowt.ie>
To: Robert Haas <robertmhaas@gmail.com>
Cc: Melanie Plageman <melanieplageman@gmail.com>, Michael Paquier <michael@paquier.xyz>, Pg Hackers <pgsql-hackers@postgresql.org>, Andres Freund <andres@anarazel.de>
Date: 2024-01-12T19:43:09Z
Lists: pgsql-hackers 
On Fri, Jan 12, 2024 at 2:32 PM Robert Haas <robertmhaas@gmail.com> wrote:
> On Fri, Jan 12, 2024 at 1:52 PM Melanie Plageman
> <melanieplageman@gmail.com> wrote:
> This analysis seems correct to me, except that "when
> lazy_scan_noprune() is called" should really say "when
> lazy_scan_noprune() is called (and returns true)", because when it
> returns false we fall through and call lazy_scan_prune() afterwards.

Now that I see your patch, I understand what Melanie must have meant.
I agree that there is a small inconsistency here, that we could well
do without.

In general I am in favor of religiously eliminating such
inconsistencies (between lazy_scan_prune and lazy_scan_noprune),
unless there is a reason not to. Not because it's necessarily
important. More because it's just too hard to be sure whether it might
matter. It's usually easier to defensively assume that it matters.

> Here's a draft patch to clean up the inconsistency here. It also gets
> rid of recordfreespace, because ISTM that recordfreespace is adding to
> the confusion here rather than helping anything.

You're using "!prunestate.has_lpdead_items" as part of your test that
sets "recordfreespace". But lazy_scan_noprune doesn't get passed a
pointer to prunestate, so clearly you'll need to detect the same
condition some other way.

-- 
Peter Geoghegan

Commits

  1. Combine FSM updates for prune and no-prune cases.

  2. Remove LVPagePruneState.

  3. Move VM update code from lazy_scan_heap() to lazy_scan_prune().

  4. Optimize vacuuming of relations with no indexes.

  5. Be more consistent about whether to update the FSM while vacuuming.

  6. Remove hastup from LVPagePruneState.

  7. Use scanned_pages to decide when to failsafe check.

  8. Simplify lazy_scan_heap's handling of scanned pages.

  9. While vacuuming a large table, update upper-level FSM data every so often.