Content-Type: text/plain; charset=utf-8; format=flowed; delsp=yes
To: =?utf-8?Q?Torsten_Z=C3=BChlsdorff?= <foo@meisterderspiele.de>
Cc: pgsql-performance@postgresql.org
Subject: Re: How to insert a bulk of data with unique-violations
 very fast
References: <hu37gm$qrl$1@news.eternal-september.org>
	<AANLkTimwLi6PkFYltlyXecsDQaI6yD3y22WLmWdgwe_R@mail.gmail.com>
	<4C07E418.80103@meisterderspiele.de>
	<AANLkTilj2xHAZMnQ-To3lod3mMPw1DBU_8EkRwTC29qb@mail.gmail.com>
	<hug2od$nu4$1@news.eternal-september.org>
	<AANLkTim36aTPinMgAQ9bUcf0AD4S1YKRrXfApHERRd7U@mail.gmail.com>
Date: Mon, 07 Jun 2010 00:35:18 +0200
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
From: "Pierre C" <lists@peufeu.com>
Message-ID: <op.vdwg04j0eorkce@apollo13>
In-Reply-To: <AANLkTim36aTPinMgAQ9bUcf0AD4S1YKRrXfApHERRd7U@mail.gmail.com>
User-Agent: Opera Mail/10.10 (Linux)

Since you have lots of data you can use parallel loading.

Split your data in several files and then do :

CREATE TEMPORARY TABLE loader1 ( ... )
COPY loader1 FROM ...

Use a TEMPORARY TABLE for this : you don't need crash-recovery since if  
something blows up, you can COPY it again... and it will be much faster  
because no WAL will be written.

If your disk is fast, COPY is cpu-bound, so if you can do 1 COPY process  
per core, and avoid writing WAL, it will scale.

This doesn't solve the other half of your problem (removing the  
duplicates) which isn't easy to parallelize, but it will make the COPY  
part a lot faster.

Note that you can have 1 core process the INSERT / removing duplicates  
while the others are handling COPY and filling temp tables, so if you  
pipeline it, you could save some time.

Does your data contain a lot of duplicates, or are they rare ? What  
percentage ?