Ignoring BRIN for HOT udpates seems broken

From: Tomas Vondra <tomas(dot)vondra(at)enterprisedb(dot)com>
To: PostgreSQL Hackers <pgsql-hackers(at)lists(dot)postgresql(dot)org>
Subject: Ignoring BRIN for HOT udpates seems broken
Date: 2022-05-28 14:50:54
Message-ID: 05ebcb44-f383-86e3-4f31-0a97a55634cf@enterprisedb.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

Hi,

while working on some BRIN stuff, I realized (my) commit 5753d4ee320b
ignoring BRIN indexes for HOT is likely broken. Consider this example:

----------------------------------------------------------------------
CREATE TABLE t (a INT) WITH (fillfactor = 10);

INSERT INTO t SELECT i
FROM generate_series(0,100000) s(i);

CREATE INDEX ON t USING BRIN (a);

UPDATE t SET a = 0 WHERE random() < 0.01;

SET enable_seqscan = off;
EXPLAIN (ANALYZE, COSTS OFF, TIMING OFF) SELECT * FROM t WHERE a = 0;

SET enable_seqscan = on;
EXPLAIN (ANALYZE, COSTS OFF, TIMING OFF) SELECT * FROM t WHERE a = 0;
----------------------------------------------------------------------

which unfortunately produces this:

QUERY PLAN
---------------------------------------------------------------
Bitmap Heap Scan on t (actual rows=23 loops=1)
Recheck Cond: (a = 0)
Rows Removed by Index Recheck: 2793
Heap Blocks: lossy=128
-> Bitmap Index Scan on t_a_idx (actual rows=1280 loops=1)
Index Cond: (a = 0)
Planning Time: 0.049 ms
Execution Time: 0.424 ms
(8 rows)

SET
QUERY PLAN
-----------------------------------------
Seq Scan on t (actual rows=995 loops=1)
Filter: (a = 0)
Rows Removed by Filter: 99006
Planning Time: 0.027 ms
Execution Time: 7.670 ms
(5 rows)

That is, the index fails to return some of the rows :-(

I don't remember the exact reasoning behind the commit, but the commit
message justifies the change like this:

There are no index pointers to individual tuples in BRIN, and the
page range summary will be updated anyway as it relies on visibility
info.

AFAICS that's a misunderstanding of how BRIN uses visibility map, or
rather does not use. In particular, bringetbitmap() does not look at the
vm at all, so it'll produce incomplete bitmap.

So it seems I made a boo boo here. Luckily, this is a PG15 commit, not a
live issue. I don't quite see if this can be salvaged - I'll think about
this a bit more, but it'll probably end with a revert.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Attachment Content-Type Size
brin.sql application/sql 423 bytes

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Daniel Gustafsson 2022-05-28 15:11:59 Re: postgres and initdb not working inside docker
Previous Message Laurenz Albe 2022-05-28 14:50:20 Re: [PATCH] Support % wildcard in extension upgrade filenames