Message-ID: <e71cf1ab-92cc-479a-b1e9-39f663867b90@gmail.com>
Date: Sat, 18 Apr 2026 21:00:00 +0300
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: eliminate xl_heap_visible to reduce WAL (and eventually set VM
 on-access)
To: Andres Freund <andres@anarazel.de>
Cc: Melanie Plageman <melanieplageman@gmail.com>,
 Tomas Vondra <tomas@vondra.me>, David Rowley <dgrowleyml@gmail.com>,
 Kirill Reshke <reshkekirill@gmail.com>, Chao Li <li.evan.chao@gmail.com>,
 Andrey Borodin <x4mmm@yandex-team.ru>, Xuneng Zhou <xunengzhou@gmail.com>,
 Robert Haas <robertmhaas@gmail.com>,
 PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>,
 Heikki Linnakangas <hlinnaka@iki.fi>
References: <2be31f17-5405-4de9-8d73-90ebc322f7d8@vondra.me>
 <CAAKRu_Ypa7-JGVR+fstDxU5Cfitk_rf5ijdaqwtoPkztursufA@mail.gmail.com>
 <CAAKRu_ZrDadxmGepBwPZ03yAKnMxwsHYn8SK9Gg7VqigLLVUWg@mail.gmail.com>
 <CAApHDvqAOeOwCKh9g0gfxWa040=Hyc7_oA=C59rjod8kXJDWyw@mail.gmail.com>
 <CAAKRu_Yt76_HdfR6DtK_wtkSNSj9=VxSV_npt+6T2R=zTzp1Pg@mail.gmail.com>
 <CAAKRu_atv6zA274m8Ysgbfn49c0NbdvHT7nXvd9kroZKnFq8Dg@mail.gmail.com>
 <CAApHDvq_R-gNXu+06GQW6w_HaEMh1pezsyiCh7GNhgh+h0UqMw@mail.gmail.com>
 <CAAKRu_YfoGTHNn0XxA+dCPj9hyO96vO4Eb+awRR6T8m22qC6ww@mail.gmail.com>
 <97529f5a-ec10-46b1-ab50-4653126c6889@gmail.com>
 <46733d68-aec0-4d09-8120-4c66b87047a4@gmail.com>
 <jfkklcxtlddx45vgx7rr27wndhkrh5umm4d2f2nhuz46lhw5ys@ohru3zfkeuww>
Content-Language: en-US
From: Alexander Lakhin <exclusion@gmail.com>
In-Reply-To: <jfkklcxtlddx45vgx7rr27wndhkrh5umm4d2f2nhuz46lhw5ys@ohru3zfkeuww>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Archived-At: 
 <https://www.postgresql.org/message-id/e71cf1ab-92cc-479a-b1e9-39f663867b90%40gmail.com>
Precedence: bulk

18.04.2026 19:25, Andres Freund wrote:
> The interesting column to show here would presumably be relallvisible.
>
> What I assume is happening is that occasionally analyze now sees enough all
> visible pages (due to on-access pruning marking the pages all visible) to
> consider the index only scan worthwhile, whereas before that wasn't (or only
> very rarely) happened.

Indeed, with c.relallvisible added, I can see:
--- .../contrib/btree_gist/expected/enum.out        2026-04-18 19:37:51.041565543 +0300
+++ .../contrib/btree_gist/results/enum.out 2026-04-18 19:40:59.077264981 +0300
@@ -88,18 +88,16 @@
  where c.relname in ('enumtmp', 'enumidx');
   relname | relpages | reltuples | autovacuum_count | autoanalyze_count | relallvisible
  ---------+----------+-----------+------------------+-------------------+---------------
- enumtmp |        3 |       595 |                0 |                 0 |             0
+ enumtmp |        3 |       595 |                0 |                 0 |             2
   enumidx |        4 |       595 | |                   |             0
  (2 rows)

  EXPLAIN (COSTS OFF)
  SELECT count(*) FROM enumtmp WHERE a >= 'g'::rainbow;
-                  QUERY PLAN
------------------------------------------------
+                   QUERY PLAN
+------------------------------------------------
   Aggregate
-   ->  Bitmap Heap Scan on enumtmp
-         Recheck Cond: (a >= 'g'::rainbow)
-         ->  Bitmap Index Scan on enumidx
-               Index Cond: (a >= 'g'::rainbow)
-(5 rows)
+   ->  Index Only Scan using enumidx on enumtmp
+         Index Cond: (a >= 'g'::rainbow)
+(3 rows)

At 378a21618~1, it stays zero.

> Maybe I'm daft, but what would prevent this from happening before? The path
> for it would be a bit more complicated, you'd have to have an autovacuum
> instead of just an analyze - but that seems possible. It might require running
> against a pre-existing install to be likely enough.

Yes, with VACUUM enumtmp; instead of ANALYZE enumtmp; the plan change is
reproduced at 378a21618~1:
@@ -88,18 +88,16 @@
  where c.relname in ('enumtmp', 'enumidx');
   relname | relpages | reltuples | autovacuum_count | autoanalyze_count | relallvisible
  ---------+----------+-----------+------------------+-------------------+---------------
- enumtmp |        3 |       595 |                0 |                 0 |             0
+ enumtmp |        3 |       595 |                0 |                 0 |             3
   enumidx |        4 |       595 | |                   |             0
  (2 rows)

  EXPLAIN (COSTS OFF)
  SELECT count(*) FROM enumtmp WHERE a >= 'g'::rainbow;
-                  QUERY PLAN
------------------------------------------------
+                   QUERY PLAN
+------------------------------------------------
   Aggregate
-   ->  Bitmap Heap Scan on enumtmp
-         Recheck Cond: (a >= 'g'::rainbow)
-         ->  Bitmap Index Scan on enumidx
-               Index Cond: (a >= 'g'::rainbow)
-(5 rows)
+   ->  Index Only Scan using enumidx on enumtmp
+         Index Cond: (a >= 'g'::rainbow)
+(3 rows)

And this diff is produced even at f7946a92 (from 2017-03-21), which added
the test case.

So, given that this is the only failure of btree_gist in two last years
at least, it looks like the probability of vacuuming the table there is
much lower than of analyzing.

>> Could you please look if this can be fixed?
> When you say fix, I assume you mean address the test instability, rather than
> actual code changes?

Sure, I didn't mean the new behavior is wrong. Probably changing that
table to temporary would work, but I wonder if there are other queries,
which plans can change due to the same reason.

Best regards,
Alexander