From: Tom Lane <tgl@sss.pgh.pa.us>
To: dddinary@163.com
cc: pgsql-bugs@lists.postgresql.org
Subject: Re: BUG #17301: SELECT gets weird result while two transactions are
 submitted concurrently
In-reply-to: <17301-61fbc4ef6a49bf89@postgresql.org>
References: <17301-61fbc4ef6a49bf89@postgresql.org>
Comments: In-reply-to PG Bug reporting form <noreply@postgresql.org>
	message dated "Sat, 27 Nov 2021 18:14:01 +0000"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <708273.1638118266.1@sss.pgh.pa.us>
Date: Sun, 28 Nov 2021 11:51:06 -0500
Message-ID: <708274.1638118266@sss.pgh.pa.us>
Archived-At: 
 <https://www.postgresql.org/message-id/708274.1638118266%40sss.pgh.pa.us>
Precedence: bulk

PG Bug reporting form <noreply@postgresql.org> writes:
> /* init */ create table t(a int primary key, b int);
> /* init */ insert into t values (1, 2), (2, 3)

> /* t1 */ begin;
> /* t1 */ set transaction isolation level repeatable read;
> /* t1 */ select * from t where a = 1;
> /* t2 */ begin;
> /* t2 */ set transaction isolation level repeatable read;
> /* t2 */ delete from t where a = 2;
> /* t2 */ commit;
> /* t1 */ update t set a = 2 where a = 1;
> /* t1 */ select * from t where a = 2;  -- [(2, 3), (2, 2)]
> /* t1 */ commit;

> The final SELECT statement gets result [(2, 3), (2, 2)], which violates the
> primary key constraint on column `a`.

This is operating as designed.  There are only three plausible behaviors
in this situation:

1. Fail t1's last SELECT (or, perhaps, its UPDATE).  You'll get that
if you use SERIALIZABLE mode.

2. In t1's last SELECT, don't show the committed-dead (2,3) row.
This violates the premise of REPEATABLE READ: t1 could see that
row at the start of its run, and it hasn't modified it, so it
should still see it.

3. In t1's last SELECT, show both rows.

We choose #3.  If you want #2, you should be using READ COMMITTED
mode, while if you want #1, you should be using SERIALIZABLE.

			regards, tom lane