Received: from localhost (postgresql.org [64.49.215.8]) by postgresql.org (Postfix) with ESMTP id 2E220476F8B; Wed, 22 Jan 2003 19:27:14 -0500 (EST) Received: from perrin.int.nxad.com (internal.ext.nxad.com [66.250.180.251]) by postgresql.org (Postfix) with ESMTP id BC38947707B; Wed, 22 Jan 2003 18:59:47 -0500 (EST) Received: by perrin.int.nxad.com (Postfix, from userid 1001) id A685E2105F; Wed, 22 Jan 2003 15:59:31 -0800 (PST) Date: Wed, 22 Jan 2003 15:59:31 -0800 From: Sean Chittenden To: Tom Lane Cc: pgsql-hackers@postgreSQL.org, pgsql-performance@postgreSQL.org Subject: Re: [PERFORM] Proposal: relaxing link between explicit JOINs and execution order Message-ID: <20030122235931.GE12075@perrin.int.nxad.com> References: <25071.1043276513@sss.pgh.pa.us> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <25071.1043276513@sss.pgh.pa.us> User-Agent: Mutt/1.4i X-PGP-Key: finger seanc@FreeBSD.org X-PGP-Fingerprint: 6CEB 1B06 BFD3 70F6 95BE 7E4D 8E85 2E0A 5F5B 3ECB X-Web-Homepage: http://sean.chittenden.org/ X-Virus-Scanned: by AMaViS new-20020517 X-Archive-Number: 200301/852 X-Sequence-Number: 34462 > There's been some recent discussion about the fact that Postgres > treats explicit JOIN syntax as constraining the actual join plan, cf > http://www.ca.postgresql.org/users-lounge/docs/7.3/postgres/explicit-joins.html > > This behavior was originally in there simply because of lack of time > to consider alternatives. I now realize that it wouldn't be hard to > get the planner to do better --- basically, preprocess_jointree just > has to be willing to fold JoinExpr-under-JoinExpr into a FromExpr > when the joins are inner joins. > > But in the meantime, some folks have found the present behavior to be > a feature rather than a bug, since it lets them control planning time > on many-table queries. If we are going to change it, I think we need > some way to accommodate both camps. [snip] > Comments? In particular, can anyone think of pithy names for these > variables? The best I'd been able to come up with is > MAX_JOIN_COLLAPSE and MAX_FROM_COLLAPSE, but neither of these > exactly sing... How about something that's runtime tunable via a SET/SHOW config var? There are some queries that I have that I haven't spent any time tuning and would love to have the planner spend its CPU thinking about it instead of mine. Setting it to 2 by default, then on my tuned queries, setting to something obscenely high so the planner won't muck with what I know is fastest (or so I think at least). I know this is a can of worms, but what about piggy backing on an Oracle notation and having an inline way of setting this inside of a comment? SELECT /* +planner:collapse_tables=12 */ .... ? ^^^^^^^ ^^^^^^^^^^^^^^^ ^^^ system variable value ::shrug:: In brainstorm mode. Anyway, a few names: auto_order_join auto_order_join_max auto_reorder_table_limit auto_collapse_join auto_collapse_num_join auto_join_threshold When I'm thinking about what this variable will do for me as a DBA, I think it will make the plan more intelligent by reordering the joins. My $0.02. -sc -- Sean Chittenden