Обсуждение: Diagnosing a massive toast file

Hi all, I have not seen this before, but one our toast tables has gotten to 15GB, the biggest table in the DB. I don't know why, is there some way I can diagnose what's causing this?

 I don't think I've ever seen a toast table even half of this size, so it's concerning.

Wells Oliver <wells.oliver@gmail.com> writes:
> Yeah, trying to figure out what actual table is clearly in need of a vacuum
> b/c of the size of that toast table.

Something like

select relname from pg_class
where reltoastrelid = 'pg_toast.pg_toast_NNN'::regclass;

(or, if you have potential duplicate relnames, select oid::regclass ...)

The mere fact that it's big does not indicate a problem, though.

            regards, tom lane

As a follow up, n_dead_tup from pg_stat_sys_tables for this TOAST table is 7447444, live tuples, 623982, and tup_del 20823469. vacuum_count is 0.

Why can't I free those rows up?

On Mon, Aug 5, 2019 at 10:00 AM Wells Oliver <wells.oliver@gmail.com> wrote:

Appreciate it, guys. I understand it being large isn't itself a problem, but relative to history and the lack of real changes, it's just strange and I'd like to better understand what is going on...

I tracked it down to a specific table, and then doing a VACUUM FULL ANALYZE on that table yields: 108765 dead row versions cannot be removed yet.

Which strikes me as odd. Any reading I can do to better understand why so many (relative to the overall table size) dead rows cannot be removed?

On Mon, Aug 5, 2019 at 9:49 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:

Wells Oliver <wells.oliver@gmail.com> writes:
> Yeah, trying to figure out what actual table is clearly in need of a vacuum
> b/c of the size of that toast table.

Something like

select relname from pg_class
where reltoastrelid = 'pg_toast.pg_toast_NNN'::regclass;

(or, if you have potential duplicate relnames, select oid::regclass ...)

The mere fact that it's big does not indicate a problem, though.

                        regards, tom lane

--

Wells Oliver
wells.oliver@gmail.com

--

Wells Oliver
wells.oliver@gmail.com

Wells Oliver <wells.oliver@gmail.com> writes:
> As a follow up, n_dead_tup from pg_stat_sys_tables for this TOAST table is
> 7447444, live tuples, 623982, and tup_del 20823469. vacuum_count is 0.

> Why can't I free those rows up?

Old open transaction somewhere (possibly a prepared transaction?).
Or a replication slot that's holding back the xmin horizon due to
not keeping up.

            regards, tom lane

Igor Neyman <ineyman@perceptron.com> writes:
> From: Wells Oliver [mailto:wells.oliver@gmail.com]
>> Do you guys have any kind of regular monitoring in place to flag users who don't politely close their connections?

> idle_in_transaction_session_timeout GUC is your friend.

Good idea, though it's only been around since 9.6.

Also, if you just want to monitor rather than impose a hard limit,
I imagine you can get Nagios to watch for this.

            regards, tom lane