Thread: Nested loop behaviour with pg_stat_activity

Nested loop behaviour with pg_stat_activity

From
Sheryl Prabhu David
Date:
Hi,

I am working on a solution to sample pg_stat_activity along the lines of what is possible in Oracle, described as follows:

In an Oracle database using a SQL like the one below, we have the ability to use, a series along the lines of Postgresqls generate_series() as seen in the 'connect by' line as outer table of a forced nested loop, and v$session(Oracles equivalent of pg_stat_activity) as the inner table, to get many samples of v$session each second.

select
        /*+ ORDERED NO_MERGE USE_NL(g t s) */
        g.*,
        t.*,
        s.*
from
        (
        select /*+ NO_MERGE */ 1 from dual connect by level <= :v_samples
        ) g,
        (
        select /*+ NO_MERGE */ hsecs from v$timer
        ) t,
        (
        select
                /*+ NO_MERGE */
                s.sql_id,
                plan_hash_value,
                sql_exec_start,
                sql_exec_id,
                event,
                state,
                wait_class,
                row_wait_obj#,
                program,
                s.module,
                s.action,
                status
        from
                v$session s
                left outer join v$sqlarea sa on s.sql_hash_value=sa.hash_value and s.sql_address=sa.address
        where
                type != 'BACKGROUND'
        ) s
order by
        sql_id,
        sql_exec_id,
        plan_hash_value,
        event,
        row_wait_obj#

Trying something similar in Postgres does not produce an equivalent result. Instead of 'Number of sample' times DIFFERENT copies of pg_stat_activity, we are seeing 'Number of sample' times SAME copy of pg_stat_activity, unlike Oracle. MVCC and Isolation guarantees for regular tables is expected to produce this kind of a result, but I was hoping pg_stat_activity being a portal into internal data structures will act similar to Oracles v$session bypassing MVCC+Isolation. I am hoping to find out if there is anyway to force Oracle type behaviour for pg_stat_activity, please help.

SQL I have tried in Postgres:

vagrant=# set enable_material = off;
SET
vagrant=# explain
with sampled as materialized (
    select
           *
    from
    (
        select
            pid,
            datname,
            usename,
            query,
            query_start,
            wait_event_type,
            wait_event,
            state
        from
            pg_stat_activity i,
            ( select * from generate_series(1,100) ) o
        offset 0
    )
    where
        state = 'active'
)
select
    query,
    query_start,
    wait_event_type,
    wait_event,
    count(*)
from
    sampled
group by
    query,
    query_start,
    wait_event_type,
    wait_event;
                                                  QUERY PLAN
---------------------------------------------------------------------------------------------------------------
 HashAggregate  (cost=330.56..331.06 rows=50 width=112)
   Group Key: sampled.query, sampled.query_start, sampled.wait_event_type, sampled.wait_event
   CTE sampled
     ->  Subquery Scan on unnamed_subquery  (cost=2.39..328.93 rows=50 width=268)
           Filter: (unnamed_subquery.state = 'active'::text)
           ->  Nested Loop  (cost=2.39..203.93 rows=10000 width=268)
                 ->  Hash Left Join  (cost=2.38..3.93 rows=100 width=268)
                       Hash Cond: (s.usesysid = u.oid)
                       ->  Hash Left Join  (cost=1.05..2.32 rows=100 width=208)
                             Hash Cond: (s.datid = d.oid)
                             ->  Function Scan on pg_stat_get_activity s  (cost=0.00..1.00 rows=100 width=148)
                             ->  Hash  (cost=1.02..1.02 rows=2 width=68)
                                   ->  Seq Scan on pg_database d  (cost=0.00..1.02 rows=2 width=68)
                       ->  Hash  (cost=1.15..1.15 rows=15 width=68)
                             ->  Seq Scan on pg_authid u  (cost=0.00..1.15 rows=15 width=68)
                 ->  Function Scan on generate_series  (cost=0.00..1.00 rows=100 width=0)
   ->  CTE Scan on sampled  (cost=0.00..1.00 rows=50 width=104)
(17 rows)

Could not get the nested loop order correct with this.

vagrant=# explain
with sampled as (
    select
           *
    from
    (
    with o as (
        select
            pid,
            datname,
            usename,
            query,
            query_start,
            wait_event_type,
            wait_event,
            state
        from
            pg_stat_activity
        offset 0
    )
    select
        o2.*
    from
        o,
        o as o1,
        o as o2
        offset 0
    )
    where
        state = 'active'
)
select
    query,
    query_start,
    wait_event_type,
    wait_event,
    count(*)
from
    sampled
group by
    query,
    query_start,
    wait_event_type,
    wait_event;
                                                              QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------
 GroupAggregate  (cost=33015.13..33139.02 rows=4889 width=112)
   Group Key: unnamed_subquery.query, unnamed_subquery.query_start, unnamed_subquery.wait_event_type, unnamed_subquery.wait_event
   ->  Sort  (cost=33015.13..33027.63 rows=5000 width=104)
         Sort Key: unnamed_subquery.query, unnamed_subquery.query_start, unnamed_subquery.wait_event_type, unnamed_subquery.wait_event
         ->  Subquery Scan on unnamed_subquery  (cost=3.93..32707.93 rows=5000 width=104)
               Filter: (unnamed_subquery.state = 'active'::text)
               ->  Nested Loop  (cost=3.93..20207.93 rows=1000000 width=268)
                     CTE o
                       ->  Hash Left Join  (cost=2.38..3.93 rows=100 width=268)
                             Hash Cond: (s.usesysid = u.oid)
                             ->  Hash Left Join  (cost=1.05..2.32 rows=100 width=208)
                                   Hash Cond: (s.datid = d.oid)
                                   ->  Function Scan on pg_stat_get_activity s  (cost=0.00..1.00 rows=100 width=148)
                                   ->  Hash  (cost=1.02..1.02 rows=2 width=68)
                                         ->  Seq Scan on pg_database d  (cost=0.00..1.02 rows=2 width=68)
                             ->  Hash  (cost=1.15..1.15 rows=15 width=68)
                                   ->  Seq Scan on pg_authid u  (cost=0.00..1.15 rows=15 width=68)
                     ->  Nested Loop  (cost=0.00..203.00 rows=10000 width=0)
                           ->  CTE Scan on o  (cost=0.00..2.00 rows=100 width=0)
                           ->  CTE Scan on o o1  (cost=0.00..2.00 rows=100 width=0)
                     ->  CTE Scan on o o2  (cost=0.00..2.00 rows=100 width=136)
(21 rows)

With this I am able to get the nested loop order correct where the alias o2 from which we are taking the results is the inner table in join order, but still we only get multiple copies of one state from pg_stat_activity.

Please suggest alternatives if any.

Regards
Prabhu David

Re: Nested loop behaviour with pg_stat_activity

From
Tom Lane
Date:
Sheryl Prabhu David <nestor.ssn@gmail.com> writes:
> In an Oracle database using a SQL like the one below, we have the ability
> to use, a series along the lines of Postgresqls generate_series() as seen
> in the 'connect by' line as outer table of a forced nested loop, and
> v$session(Oracles equivalent of pg_stat_activity) as the inner table, to
> get many samples of v$session each second.

That's ... impressively ill-defined.  How do you know you are getting
consistent "samples" at all?

> Trying something similar in Postgres does not produce an equivalent result.
> Instead of 'Number of sample' times *DIFFERENT* copies of pg_stat_activity,
> we are seeing 'Number of sample' times *SAME* copy of pg_stat_activity,
> unlike Oracle. MVCC and Isolation guarantees for regular tables is expected
> to produce this kind of a result, but I was hoping pg_stat_activity being a
> portal into internal data structures will act similar to Oracles v$session
> bypassing MVCC+Isolation. I am hoping to find out if there is anyway to
> force Oracle type behaviour for pg_stat_activity, please help.

A Postgres session captures a snapshot of pg_stat_activity on first
reference, and holds it until end of transaction or you call
pg_stat_clear_snapshot().  Without this behavior you would get total
garbage from queries as the data changes under you, especially so
from join queries which may require multiple scans of the input.

I don't think there's any way to precisely duplicate what you describe
doing in Postgres, but you can easily get a similar result by
alternating "SELECT pg_stat_clear_snapshot()" with selects from
pg_stat_activity.

            regards, tom lane



Re: Nested loop behaviour with pg_stat_activity

From
Sheryl Prabhu David
Date:
Thanks a lot Tom, Postgres community truly is impressive, I never thought I will get a response this soon if at all.

In Oracle with v$session being the inner table of a nested loop for each scan we get different copies of it even as the data changes under us, thats the hack someone figured out, and it gives us tons of samples - in the order of ~500 per second with a single SQL run. 

In Postgres the reason I mention the behaviour is different is that the count I get with the outer group by for each connected client is one record with a count of 2000 if I use 2000 records from generate_series. Oracles behaviour would be multiple records each with a different wait event adding up to 2000.

As you describe this may not be possible with Postgres. I will try SELECT pg_stat_clear_snapshot() but not sure if that can be done is a single SQL statement. Will let you know if it works.

Thanks a lot again.

Regards
Prabhu David

On Wed, Mar 13, 2024 at 8:00 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Sheryl Prabhu David <nestor.ssn@gmail.com> writes:
> In an Oracle database using a SQL like the one below, we have the ability
> to use, a series along the lines of Postgresqls generate_series() as seen
> in the 'connect by' line as outer table of a forced nested loop, and
> v$session(Oracles equivalent of pg_stat_activity) as the inner table, to
> get many samples of v$session each second.

That's ... impressively ill-defined.  How do you know you are getting
consistent "samples" at all?

> Trying something similar in Postgres does not produce an equivalent result.
> Instead of 'Number of sample' times *DIFFERENT* copies of pg_stat_activity,
> we are seeing 'Number of sample' times *SAME* copy of pg_stat_activity,
> unlike Oracle. MVCC and Isolation guarantees for regular tables is expected
> to produce this kind of a result, but I was hoping pg_stat_activity being a
> portal into internal data structures will act similar to Oracles v$session
> bypassing MVCC+Isolation. I am hoping to find out if there is anyway to
> force Oracle type behaviour for pg_stat_activity, please help.

A Postgres session captures a snapshot of pg_stat_activity on first
reference, and holds it until end of transaction or you call
pg_stat_clear_snapshot().  Without this behavior you would get total
garbage from queries as the data changes under you, especially so
from join queries which may require multiple scans of the input.

I don't think there's any way to precisely duplicate what you describe
doing in Postgres, but you can easily get a similar result by
alternating "SELECT pg_stat_clear_snapshot()" with selects from
pg_stat_activity.

                        regards, tom lane