Message-ID: <50A4497D.80902@dunslane.net>
Date: Wed, 14 Nov 2012 20:46:37 -0500
From: Andrew Dunstan <andrew@dunslane.net>
User-Agent: Mozilla/5.0 (X11; Linux x86_64;
	rv:15.0) Gecko/20120827 Thunderbird/15.0
MIME-Version: 1.0
To: Craig Ringer <craig@2ndQuadrant.com>
CC: Tom Lane <tgl@sss.pgh.pa.us>, David Greco <David_Greco@harte-hanks.com>,
	"pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: SOLVED - RE: Poor performance using CTE
References: 
 <187F6C10D2931A4386EE8E58E13857F61291D897@BY2PRD0811MB415.namprd08.prod.outlook.com>
	<20462.1352910543@sss.pgh.pa.us> <50A442A8.10803@2ndQuadrant.com>
In-Reply-To: <50A442A8.10803@2ndQuadrant.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit


On 11/14/2012 08:17 PM, Craig Ringer wrote:
> On 11/15/2012 12:29 AM, Tom Lane wrote:
>> David Greco <David_Greco@harte-hanks.com> writes:
>>> Thanks, that did the trick. Though I'm still not clear as to why.
>> PG treats WITH as an optimization fence --- the WITH query will be
>> executed pretty much as-is.  It may be that Oracle flattens the query
>> somehow; though if you're using black-box functions in both cases,
>> it's not obvious where the optimizer could get any purchase that way.
>>
> I was looking through the latest spec drafts I have access to and
> couldn't find any reference to Pg's optimisation-fence-for-CTEs
> behaviour being required by the standard, though I've repeatedly seen it
> said that there is such a requirement.
>
> Do you know where it's specified?
>
> All I can see is that the optimised result must have the same effect as
> the original. That'd mean that wCTEs and CTE terms that use VOLATILE
> functions or functions with side-effects couldn't be optimised into
> other queries. Simple CTEs could be, though, and there are times I've
> really wished I could use a CTE but I've had to use a set-returning
> subquery to get reasonable plans.


It cuts both ways. I have used CTEs a LOT precisely because this 
behaviour lets me get better plans. Without that I'll be back to using 
the "offset 0" hack.

cheers

andrew