To: Richard Huxton <dev@archonet.com>
cc: Tony Cebzanov <tonyceb@andrew.cmu.edu>, pgsql-sql@postgresql.org
Subject: Re: Avoiding cycles in a directed graph 
In-reply-to: <4BA002DC.1050700@archonet.com> 
References: <4B9FDFDE.5030005@andrew.cmu.edu> <4B9FE752.4080709@archonet.com>
	<14113.1268771643@sss.pgh.pa.us> <4B9FEEE8.1060404@andrew.cmu.edu>
	<14550.1268773752@sss.pgh.pa.us> <4BA002DC.1050700@archonet.com>
Comments: In-reply-to Richard Huxton <dev@archonet.com>
	message dated "Tue, 16 Mar 2010 22:14:52 -0000"
Date: Tue, 16 Mar 2010 18:22:12 -0400
Message-ID: <15827.1268778132@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Richard Huxton <dev@archonet.com> writes:
> On 16/03/10 21:09, Tom Lane wrote:
>> If you don't expect this to be common, maybe you could fix the
>> concurrency issue by taking a table-wide lock that locks out
>> other writers.

> Surely SELECT FOR UPDATE on the parents would be sufficient? If there's 
> no overlap between (currently non-cyclic) graphs being altered then 
> there can't be any conflict.

Um, what if the cycle is being formed from whole cloth?  For instance
T1 inserts an edge A->B while T2 is inserting B->A.  There are no
pre-existing rows to lock, but there will still be a cycle after they
both commit.

Also it seems pretty deadlock-prone if there are multiple existing rows
to try to lock.  Perhaps you could work around the risk by locking those
rows one at a time in an application-defined ordering ... but I'm afraid
the performance would be poor, unless the connected graphs are always
very small.

On the whole I think Tony's better off with a KISS approach, ie just
lock the whole table against other writers.

			regards, tom lane