Обсуждение: [PATCH] pg_stat_statements dealloc field ignores manual deallocation

Hi,

This is a proposal of a patch for pg_stat_statements extension. It
corrects deallocation events accounting.

Since 2e0fedf there is a view pg_stat_statements_info is available in
pg_stat_statements extension. It has a dealloc field, that should be a
counter of deallocation events happened.
Right now it accounts only automatic deallocation events, happened when
we need a place for a new statement, but manual deallocation events
caused by pg_stat_statements_reset() function for some subset of
collected statements is not accounted.
My opinion is that manual deallocation is a deallocation too and it
must be accounted in dealloc field of pg_stat_statements_info view.

Let's see how it happens:

postgres=# select pg_stat_statements_reset();
postgres=# select 1;
 ?column?
----------
        1
(1 row)
postgres=# select dealloc from pg_stat_statements_info ;
 dealloc
---------
       0
(1 row)

postgres=# select pg_stat_statements_reset(userid,dbid,queryid)
postgres-# from pg_stat_statements where query = 'select $1';
 pg_stat_statements_reset
--------------------------

(1 row)

postgres=# select dealloc from pg_stat_statements_info ;
 dealloc
---------
       0  -- Here must be a one now, as deallocation happened
(1 row)

This patch adds accounting of manual deallocation events.

-- 
Andrei Zubkov
Postgres Professional
The Russian Postgres Company

On Fri, Mar 19, 2021 at 05:08:45PM +0300, Андрей Зубков wrote:
> 
> Since 2e0fedf there is a view pg_stat_statements_info is available in
> pg_stat_statements extension. It has a dealloc field, that should be a
> counter of deallocation events happened.
> Right now it accounts only automatic deallocation events, happened when
> we need a place for a new statement,

Yes, and that behavior is documented:

dealloc bigint

Total number of times pg_stat_statements entries about the least-executed
statements were deallocated because more distinct statements than
pg_stat_statements.max were observed

> but manual deallocation events
> caused by pg_stat_statements_reset() function for some subset of
> collected statements is not accounted.
> My opinion is that manual deallocation is a deallocation too and it
> must be accounted in dealloc field of pg_stat_statements_info view.

I disagree.  The point of that field is to help users configuring
pg_stat_statements.max, as evictions have a huge overhead in many workloads.

If users remove entries for some reasons, we don't have to give the impression
that pg_stat_statements.max is too low and that it should be increased,
especially since it requires a restart.

On Fri, 2021-03-19 at 22:15 +0800, Julien Rouhaud wrote:
> I disagree.  The point of that field is to help users configuring
> pg_stat_statements.max, as evictions have a huge overhead in many
> workloads.
> 
> If users remove entries for some reasons, we don't have to give the
> impression
> that pg_stat_statements.max is too low and that it should be
> increased,
> especially since it requires a restart.
> 
Ok.
But when we are collecting aggregated statistics on pg_stat_statememts
periodically it would be great to know about every deallocation
happened. Maybe we need to add another counter for manual deallocations
tracking?