Date: Wed, 14 Nov 2007 22:26:02 -0500
From: Jean-David Beyer <jeandavid8@verizon.net>
Subject: Re: Curious about dead rows.
In-reply-to: <16999.1195080384@sss.pgh.pa.us>
To: pgsql-performance@postgresql.org
Message-id: <473BBC4A.20304@verizon.net>
Organization: Institute for Regimented Whimsey
MIME-version: 1.0
Content-type: text/plain; charset=ISO-8859-1
Content-transfer-encoding: 7bit
References: <4735DBDB.5050208@verizon.net> <16627.1194714491@sss.pgh.pa.us>
	<4735FA9F.1080805@verizon.net>
	<b42b73150711131628q1469df3el6f557e10de07a05@mail.gmail.com>
	<473A5CC0.3060003@verizon.net>
	<b42b73150711131936l65b97b9ya7953383673b2710@mail.gmail.com>
	<473AE63D.7090305@verizon.net> <20071114152249.GA16041@crankycanuck.ca>
	<473B2A1F.30804@verizon.net> <473B6C9D.1020804@pws.com.au>
	<16999.1195080384@sss.pgh.pa.us>
User-Agent: Thunderbird 1.5.0.9 (X11/20070317)

Tom Lane wrote:
> Russell Smith <mr-russ@pws.com.au> writes:
>> It is possible that analyze is not getting the number of dead rows
>> right?
> 
> Hah, I think you are on to something.  ANALYZE is telling the truth about
> how many "dead" rows it saw, but its notion of "dead" is "not good 
> according to SnapshotNow".  Thus, rows inserted by a not-yet-committed 
> transaction would be counted as dead.  So if these are background 
> auto-analyzes being done in parallel with inserting transactions that run
> for awhile,

They are.

> seeing a few not-yet-committed rows would be unsurprising.

That is a very interesting possibility. I can see that it is certainly a
possible explanation, since my insert transactions take between 0.04 to 0.1
minutes (sorry, decimal stopwatch) of real time, typically putting 1700 rows
into about a half dozen tables. And the ANALYZE is whatever autovacuum
chooses to do. So if new not-yet-committed rows are considered dead, that
would be a sufficient explanation.

So I am, retroactively, unsurprised.

> I wonder if that is worth fixing?  I'm not especially concerned about the
> cosmetic aspect of it, but if we mistakenly launch an autovacuum on the
> strength of an inflated estimate of dead rows, that could be costly.
> 
Well, since I was more interested in the explanation than in the fixing, in
that sense I do not care if it is fixed or not. While it may create a slight
slowdown (if it is an error), the applications run "fast enough."

I would not even get the fix until Red Hat get around to putting it in (I
run postgresql-8.1.9-1.el5 that is in their RHEL5 distribution), that
probably will not be until RHEL6 and the soonest, and I will probably skip
that one and wait until RHEL7 comes out in about 3 years.

But somewhere perhaps a reminder of this should be placed where someone like
me would find it, so we would not have to go through this again for someone
else.

-- 
  .~.  Jean-David Beyer          Registered Linux User 85642.
  /V\  PGP-Key: 9A2FC99A         Registered Machine   241939.
 /( )\ Shrewsbury, New Jersey    http://counter.li.org
 ^^-^^ 22:05:01 up 22 days, 15:23, 0 users, load average: 4.16, 4.22, 4.10