Re: Worse performance with higher work_mem?

On Jan 13, 2020, at 6:34 PM, Dilip Kumar <dilipbalaut@gmail.com> wrote:

On Tue, Jan 14, 2020 at 5:29 AM Israel Brewster <ijbrewster@alaska.edu> wrote:

I was working on diagnosing a “slow” (about 6 second run time) query:

SELECT
           to_char(bucket,'YYYY-MM-DD"T"HH24:MI:SS') as dates,
           x_tilt,
           y_tilt,
           rot_x,
           rot_y,
           date_part('epoch', bucket) as timestamps,
           temp
       FROM
           (SELECT
             time_bucket('1 week', read_time) as bucket,
             avg(tilt_x::float) as x_tilt,
             avg(tilt_y::float) as y_tilt,
             avg(rot_x::float) as rot_x,
             avg(rot_y::float) as rot_y,
             avg(temperature::float) as temp
           FROM tilt_data
           WHERE station='c08883c0-fbe5-11e9-bd6e-aec49259cebb'
           AND read_time::date<='2020-01-13'::date
           GROUP BY bucket) s1
       ORDER BY bucket;

In looking at the explain analyze output, I noticed that it had an “external merge Disk” sort going on, accounting for about 1 second of the runtime (explain analyze output here: https://explain.depesz.com/s/jx0q). Since the machine has plenty of RAM available, I went ahead and increased the work_mem parameter. Whereupon the query plan got much simpler, and performance of said query completely tanked, increasing to about 15.5 seconds runtime (https://explain.depesz.com/s/Kl0S), most of which was in a HashAggregate.

I am running PostgreSQL 11.6 on a machine with 128GB of ram (so, like I said, plenty of RAM)

How can I fix this? Thanks.

I have noticed that after increasing the work_mem your plan has
switched from a parallel plan to a non-parallel plan.  Basically,
earlier it was getting executed with 3 workers.  And, after it becomes
non-parallel plan execution time is 3x.  For the analysis can we just
reduce the value of parallel_tuple_cost and parallel_setup_cost and
see how it behaves?

-- 
Regards,
Dilip Kumar
EnterpriseDB: http://www.enterprisedb.com