Re: [PERFORM] performance drop after upgrade (9.6 > 10)

2017-10-24 Thread Johannes Graën
On 2017-10-24 17:18, Justin Pryzby wrote:
> You could (re)install PG96 alongside PG10 and run a copy of the DB (even from
> your homedir, or on a difference server) and pg_dump |pg_restore the relevant
> tables (just be sure to specify the alternate host/port/user/etc as needed for
> the restore invocation).

I considered that but it is far too expensive just for getting the old
query plan. The database is more than 1 TB big and replaying it from a
dump to another server took us several days, primarily due to the heavy
use of materialized views that are calculated over all rows of some
large tables. As long as there is no safe pg_downgrade --link I'd rather
keep trying to improve query performance on the current version.


-- 
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance


Re: [PERFORM] performance drop after upgrade (9.6 > 10)

2017-10-24 Thread johannes graën
Hi Pavel, *,

you were right with ANALYZing the DB first. However, even after doing
so, I frequently see Seq Scans where an index was used before. This
usually cooccurs with parallelization and looked different before
upgrading to 10. I can provide an example for 10 [1], but I cannot
generate a query plan for 9.6 anymore.

Any ideas what makes the new version more seqscanny?


[1] https://explain.depesz.com/s/gXD3


-- 
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance


Re: [PERFORM] performance drop after upgrade (9.6 > 10)

2017-10-11 Thread johannes graën
On Wed, Oct 11, 2017 at 1:11 PM, Pavel Stehule  wrote:
> have you fresh statistics? After upgrade is necessary to run ANALYZE command

Yes, that was missing indeed. I did ANALYZE but apparently on all
databases but this one. I could have guessed that
1,098,956,679,131,935,754,413,282,631,696,252,928 is not a reasonable
cost value.

Thanks, Pavel.

Best
  Johannes


-- 
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance


[PERFORM] performance drop after upgrade (9.6 > 10)

2017-10-11 Thread johannes graën
Hi,

I wrote a query that joins several tables usually returning less than
1000 rows, groups them and generates a JSON object of the result. In
9.6 is was a question of milliseconds for that query to return the
requested data. Now, after upgrading to 10, the query never returns -
at least it hasn't returned in the last hour.

To see what happens, I requested the query plan [1]. It looks complex
and shows a lot of parallelization. I don't have the query plan from
9.6, but I remember it being considerably simpler.

Can anyone have a guess what altered the performance here so
dramatically? Is there a way to disable new parallelization features
just for this query to see if it makes any difference?

Best
  Johannes


[1] https://explain.depesz.com/s/xsPP


-- 
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance