From: Tom Lane <tgl@sss.pgh.pa.us>
To: Marcin Gozdalik <gozdal@gmail.com>
cc: pgsql-performance@postgresql.org
Subject: Re: Very slow "bloat query"
In-reply-to: 
 <CADu1mRN1PPXO0p2HUF4NkrLSLmdiXXcafm5B6r2kk7Tawd1UoQ@mail.gmail.com>
References: 
 <CADu1mRN1PPXO0p2HUF4NkrLSLmdiXXcafm5B6r2kk7Tawd1UoQ@mail.gmail.com>
Comments: In-reply-to Marcin Gozdalik <gozdal@gmail.com>
	message dated "Fri, 14 May 2021 11:06:33 -0000"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <1884330.1621004642.1@sss.pgh.pa.us>
Date: Fri, 14 May 2021 11:04:02 -0400
Message-ID: <1884331.1621004642@sss.pgh.pa.us>
Archived-At: 
 <https://www.postgresql.org/message-id/1884331.1621004642%40sss.pgh.pa.us>
Precedence: bulk

Marcin Gozdalik <gozdal@gmail.com> writes:
> I have traced the problem to the bloated `pg_class` (the irony: `pgmetrics`
> does not collect bloat on `pg_catalog`):
> `vacuum (full, analyze, verbose) pg_class;`
> ```
> INFO:  vacuuming "pg_catalog.pg_class"
> INFO:  "pg_class": found 1 removable, 7430805 nonremovable row versions in
> 158870 pages
> DETAIL:  7429943 dead row versions cannot be removed yet.

Ugh.  It's understandable that having a lot of temp-table traffic
would result in the creation of lots of dead rows in pg_class.
The question to be asking is why aren't they vacuumable?  You
must have a longstanding open transaction somewhere (perhaps
a forgotten prepared transaction?) that is holding back the
global xmin horizon.  Closing that out and then doing another
manual VACUUM FULL should help.

			regards, tom lane