X-Original-To: pgsql-performance-postgresql.org@localhost.postgresql.org Received: from localhost (unknown [64.117.224.130]) by svr1.postgresql.org (Postfix) with ESMTP id F36BBD1B4EB for ; Sun, 7 Sep 2003 23:18:35 +0000 (GMT) Received: from svr1.postgresql.org ([64.117.224.193]) by localhost (neptune.hub.org [64.117.224.130]) (amavisd-new, port 10024) with ESMTP id 86785-10 for ; Sun, 7 Sep 2003 20:18:06 -0300 (ADT) Received: from smtp.istop.com (dci.doncaster.on.ca [66.11.168.194]) by svr1.postgresql.org (Postfix) with ESMTP id D7E7AD1B516 for ; Sun, 7 Sep 2003 20:18:02 -0300 (ADT) Received: from stark.dyndns.tv (gsstark.mtl.istop.com [66.11.160.162]) by smtp.istop.com (Postfix) with ESMTP id 5BFF736A9B; Sun, 7 Sep 2003 19:18:02 -0400 (EDT) Received: from localhost ([127.0.0.1] helo=stark.dyndns.tv ident=foobar) by stark.dyndns.tv with smtp (Exim 3.36 #1 (Debian)) id 19w8nC-0003OM-00; Sun, 07 Sep 2003 19:18:02 -0400 To: Mary Edie Meredith Cc: Tom Lane , pgsql-performance , osdldbt-general Subject: Re: [GENERAL] how to get accurate values in pg_statistic References: <1062697270.6854.3475.camel@ibm-e.pdx.osdl.net> <11537.1062717375@sss.pgh.pa.us> <1062783865.6854.4780.camel@ibm-e.pdx.osdl.net> In-Reply-To: <1062783865.6854.4780.camel@ibm-e.pdx.osdl.net> From: Greg Stark Organization: The Emacs Conspiracy; member since 1992 Date: 07 Sep 2003 19:18:01 -0400 Message-ID: <87fzj86vdi.fsf@stark.dyndns.tv> Lines: 21 User-Agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.3 MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii X-Virus-Scanned: by amavisd-new at postgresql.org X-Archive-Number: 200309/116 X-Sequence-Number: 3482 Mary Edie Meredith writes: > We ran additional tests with default_statistics_target set to 1000 (the > max I believe). The plans are the same over the different runs, but the > pg_statistics table has different cost values. The performance results > of the runs are consistent (we would expect this with the same plans). > The resulting performance metrics are similar to the best plans we see > using the default histogram size (good news). Hm, would it be possible to do a binary search and find the target at which you start getting consistent plans? Perhaps the default of 10 is simply way too small and should be raised? Obviously this would depend on the data model, but I suspect if your aim is for the benchmark data to be representative of typical data models, which scares me into thinking perhaps users are seeing similarly unpredictably variable performance. -- greg