To: Samuel Tardieu <sam@rfc1149.net>
Cc: pgsql-sql@postgresql.org
Subject: Re: Question about isolation 
In-reply-to: <871xpj928m.fsf@beeblebrox.enst.fr> 
References: <Pine.BSO.4.44.0401281507300.30592-100000@detroit.arbor.net>
	<Pine.BSO.4.44.0401281534400.30592-100000@detroit.arbor.net>
	<871xpj928m.fsf@beeblebrox.enst.fr>
Comments: In-reply-to Samuel Tardieu <sam@rfc1149.net>
	message dated "Wed, 28 Jan 2004 22:45:13 +0100"
Date: Wed, 28 Jan 2004 18:03:25 -0500
Message-ID: <7527.1075331005@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Samuel Tardieu <sam@rfc1149.net> writes:
>>> Do I have the guarantee that, in any event, rows deleted from
>>> table t by the delete won't reappear in the select result?
>> 
>> i do not think you have that guarantee in READ COMMITTED mode
>> because there is a slight possibility another backend sneaked a
>> committed insert in between the delete and select
>> statement.

> Yes. But the possible effect your describe (insertion of new rows
> after the DELETE statement and before the SELECT) matches accurately
> the symptoms we are observing.

Hmm.  I think you need to look closer.  AFAIR the READ COMMITTED
behavior is only an issue if you give the commands interactively
from the client.  Inside a plpgsql function we do not do
SetQuerySnapshot() and therefore the snapshot of other transactions'
effects does not advance.  So I think the coding should be safe
... at the moment.  (A number of people think the lack of
SetQuerySnapshot inside functions is a bug; so the behavior might
change in future.)

Using SERIALIZABLE mode would probably make your code more future-proof,
but if you are presently seeing failures, there's some other effect
involved here.

			regards, tom lane