Message-ID: <4ED6BE16.5090706@ringerc.id.au>
Date: Thu, 01 Dec 2011 07:36:54 +0800
From: Craig Ringer <ringerc@ringerc.id.au>
User-Agent: Mozilla/5.0 (X11; Linux x86_64;
	rv:7.0.1) Gecko/20110930 Thunderbird/7.0.1
MIME-Version: 1.0
To: "Nicholson, Brad (Toronto, ON, CA)" <bnicholson@hp.com>
CC: Rob Richardson <RDRichardson@rad-con.com>,
	"pgsql-admin@postgresql.org" <pgsql-admin@postgresql.org>
Subject: Re: transaction error handling
References: 
 <CAANiQEm92w70rapQhitV1qL0tpkNRo5qjpwo2fdg+iZ-AURMCg@mail.gmail.com>
	<28131.1322067903@sss.pgh.pa.us>
	<232B5217AD58584C87019E8933556D11036BDE58EB@redmx2.esri.com>
	<67D108EDFAD3C148A593E6ED7DCB4BBDEDDBAE@RADCONWIN2K8PDC.radcon.local>
	<EC55DC235432104F8255702A8D7344D9256FF7EF@G9W0741.americas.hpqcorp.net>
In-Reply-To: 
 <EC55DC235432104F8255702A8D7344D9256FF7EF@G9W0741.americas.hpqcorp.net>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

On 11/30/2011 09:19 PM, Nicholson, Brad (Toronto, ON, CA) wrote:
> This functionality is something that Postgres can do today. We expose 
> the ability to do this with explicit savepoints. The difference is 
> that Oracle allows you to set it on a per transaction basis (I 
> believe) and it will behave this way for all statements in the 
> transaction, where as we need to do it explicitly. Looking through the 
> archives there does seem to be a performance problem on commit in 
> Postgres if you issue a lot of savepoints (there were discussions of a 
> fix but I am not sure the status of this).
Savepoint performance has had several improvements over time. Back in 
8.1 when I got started developing against Pg seriously I was having 
*MASSIVE* performance issues with PL/PgSQL exception blocks (which use 
savepoints) in loops; these days it's perfect.

To make automatic savepoints viable, Pg would need to be able to 
completely forget a savepoint once it's been released, so there's no 
ongoing cost. That way a transaction would only need two savepoints at 
any particular point in time. My understanding is that it's not there 
yet; AFAIK released savepoints still have a non-trivial cost that would 
add up if someone was using automatic savepoints in (say) a 10,000 
INSERT transaction.

--
Craig Ringer