To: Florian Weimer <fweimer@bfk.de>
cc: pgsql-sql@postgresql.org
Subject: Re: INSERT/UPDATEs cycles and lack of phantom locking 
In-reply-to: <82zmf60zf4.fsf@mid.bfk.de> 
References: <82zmf60zf4.fsf@mid.bfk.de>
Comments: In-reply-to Florian Weimer <fweimer@bfk.de>
	message dated "Wed, 19 Jul 2006 10:30:39 +0200"
Date: Wed, 19 Jul 2006 10:26:18 -0400
Message-ID: <25634.1153319178@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Florian Weimer <fweimer@bfk.de> writes:
> BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED; -- SERIALIZABLE
> ...
> -- SAVEPOINT tmp_created;  -- (see below)

> CREATE TEMPORARY TABLE tmp2 AS SELECT tmp.key, new_val, real.val AS old_val
>   FROM tmp LEFT OUTER JOIN real ON tmp.key = real.key;
> UPDATE real SET val = new_val + old_val FROM tmp2
>   WHERE old_val IS NOT NULL AND tmp2.key = real.key;
> INSERT INTO real SELECT key, new_val FROM tmp2 WHERE old_val IS NULL;

> If this is run concurrently, the INSERT may fail.  In this case, I
> rerun the transaction.  Actually, I want to rollback to the
> tmp_created checkpoint, but I don't think this will pick up the new
> rows in the "real" table, and the INSERT will fail again.

Why do you think that?  If you're running in READ COMMITTED mode then
each statement takes a new snapshot.

			regards, tom lane