Обсуждение: Performance Issues

On Tue, Apr 23, 2002 at 02:16:46PM -0500,
  Shaun Grannis <shaun_grannis@hotmail.com> wrote:
> Hi,
>
>
>
> I've found some performance issues with Postgres that I'm hoping people on this list can help resolve.  We're working
witha 65 million record table that includes year of birth (data type INT). To count the frequency of dates in the
table,it takes 2 hours 26 minutes to execute. (There's an approximately 100-year range of dates in the 65 million
records).

I believe that issue is the postgres doesn't know how many different values
to expect, so that rather than set up buckets and add one to the appropiate
bucket when you encouter a value (which is what your perl script does),
it sorts all of the like values together and counts them up one group at
a time. The sort is probably what is killing your performance.

There has been some discussion on this topic on one of the postgres
lists in the last couple of days.

> # EXPLAIN SELECT ln, count(1) FROM data_table GROUP BY ln;
>
> NOTICE:  QUERY PLAN:
>
> Aggregate (cost=19538149.27..19864263.69 rows=6522288 width=19)
>
>   ->  Group (cost=19538149.27..19701206.48 rows=65222884 width=19)
>
>        ->  Sort (cost=19538149.27..19538149.27 rows=65222884 width=19)
>
>             ->  Seq Scan on data_table (cost=0.00..2324610.84 rows=65222884 width=19)
>
>
>
> The last name (ln) and the year of birth (yb) is indexed, but that
> shouldn't matter because it's doing a sequential scan, correct?  Am I
> running into the limitations of Postgres?  We'd like to eventually get

I didn't see anything about your settings in postgresql.conf, but
increasing the sort_mem parameter may help that really expensive sort
step.  I think the desired fix for this would probably be the TODO entry
on hash based aggregates but that's still in the future...

Le Mardi 23 Avril 2002 21:16, Shaun Grannis a écrit :
> Thanks for you suggestions.

Dear Shaun,

The only solution I see is to create a pseudo-count function.
For this, you need to :

- add fire_trigger (timestamp) field on data_table with index.

- create a datefreq_table table which will be used
  as a pseudo-counter with two fields :
  datefreq_date (int4) and datefreq_count (int8).
  Both fields should be indexed.

- create a trigger on after update or insert of data_table
  to increase datefreq_count by 1.

- create a trigger on after delete of data_table
  to decrease date_freq_count by 1.

- create the needed records in datefreq_table (100 as you said).

To inizialize the system, enter:
UPDATE * FROM data_table
SET fire_trigger ='now'
WHERE fire_trigger IS NULL;
LIMIT 1000

This will inizilize the system for 10000 records. I don't know how much time
it will take. Probably 10 seconds on your hardware. Mesure the time needed to
inizialize the whole system.

Once inizialized, the system will give immediate answers using the
pseudo-counter table.

SELECT datefreq_count
FROM datefreq_table
WHERE datefreq_date = foo;

The only drawback of such a system is that you will not be able to DROP or
INSERT or even UPDATE data as fast as if there was no trigger. it will only
be a problem if your data changes very often.

Even on Oracle or DB2, you would have to use data-mining features to be able
to manage such amounts of data. Here, on PostgreSQL, you can do it manually
using PLpgSQL.

Cheers,
Jean-Michel