Received: from sss.sss.pgh.pa.us (sss.pgh.pa.us [209.114.166.2]) by hub.org (8.9.3/8.9.3) with ESMTP id LAA74722 for ; Wed, 22 Sep 1999 11:14:06 -0400 (EDT) (envelope-from tgl@sss.pgh.pa.us) Received: from sss.sss.pgh.pa.us (localhost [127.0.0.1]) by sss.sss.pgh.pa.us (8.9.1/8.9.1) with ESMTP id LAA00225; Wed, 22 Sep 1999 11:13:31 -0400 (EDT) To: Stuart Rison cc: pgsql-sql Subject: Re: [SQL] [GENERAL] UPDATE feature or bug? In-reply-to: Your message of Wed, 22 Sep 1999 13:35:51 +0100 (BST) Date: Wed, 22 Sep 1999 11:13:31 -0400 Message-ID: <222.938013211@sss.pgh.pa.us> From: Tom Lane Stuart Rison writes: > Now I've sussed this out as the update must be dealing with rows where b=4 > first, trying to UPDATE them to b=3 and failing because these are already > 'occupied'. > Is this a bug? is this a feature? is there a way of telling the UPDATE > what order to UPDATE in? It's a bug, but I doubt it's likely to get fixed any time soon, since unique indexes would get a LOT slower if they did this "right". As you say, the update must be hitting some of the b=4 rows first. At that point, the system has no idea that the existing b=3 rows are going to be invalidated in the same transaction, so its uniqueness check fails. To do this right, the uniqueness check would have to be applied at the end of the transaction, which would mean remembering every tuple that's been inserted/updated in the current transaction and going back to check its uniqueness. There is a running discussion on the hackers list about implementing referential integrity constraints, which have this same problem that you can't really apply (some kinds of) checks until you see the entire transaction results. It'll get done but it's not simple. I'm not sure whether anyone is thinking of applying that mechanism to unique- index checks... regards, tom lane