Обсуждение: Odd Row Estimates in Query Plan (rows=75)

On 08/15/2018 12:31 PM, Don Seiler wrote:
> PostgreSQL 9.6.6 on CentOS.
> 
> We have a report query that has gone from maybe a few seconds to run to 
> a few minutes to run since mid-July. Looking at the output of EXPLAIN 
> ANALYZE, the row count estimates are way off, even though this table was 
> just analyzed a day or so ago. What's more bizarre to me is that the row 
> count esimate is *always* 75 for every node of the plan, where the 
> actual rows is in the hundreds or thousands. This table is one of the 
> busiest tables in our production database (many inserts and updates). It 
> is autovacuumed and autoanalyzed a few times per week, although I'm 
> looking to change it to a nightly manual schedule to avoid daytime 
> autovacuums.
> 
> Hash Join  (cost=1869142.34..1869146.15 rows=75 width=88) (actual 
> time=179877.869..179878.011 rows=759 loops=1)
>     Hash Cond: (stores.pkey = lt.store_pkey)
>     Buffers: shared hit=1654593 read=331897 dirtied=249
>     ->  Seq Scan on stores  (cost=0.00..2.77 rows=77 width=22) (actual 
> time=0.007..0.023 rows=78 loops=1)
>           Buffers: shared hit=2
>     ->  Hash  (cost=1869141.40..1869141.40 rows=75 width=50) (actual 
> time=179877.847..179877.847 rows=759 loops=1)
>           Buckets: 1024  Batches: 1  Memory Usage: 73kB
>           Buffers: shared hit=1654591 read=331897 dirtied=249
>           ->  Subquery Scan on lt  (cost=1869138.59..1869141.40 rows=75 
> width=50) (actual time=179875.976..179877.697 rows=759 loops=1)
>                 Buffers: shared hit=1654591 read=331897 dirtied=249
>                 ->  GroupAggregate  (cost=1869138.59..1869140.65 rows=75 
> width=50) (actual time=179875.976..179877.606 rows=759 loops=1)
>                       Group Key: lts.store_pkey, lts.owner, 
> (date_trunc('minute'::text, lts.date_gifted))
>                       Filter: (count(*) IS NOT NULL)
>                       Buffers: shared hit=1654591 read=331897 dirtied=249
>                       ->  Sort  (cost=1869138.59..1869138.78 rows=75 
> width=42) (actual time=179875.961..179876.470 rows=6731 loops=1)
>                             Sort Key: lts.store_pkey, 
> lts.entry_source_owner, (date_trunc('minute'::text, lts.date_gifted))
>                             Sort Method: quicksort  Memory: 757kB
>                             Buffers: shared hit=1654591 read=331897 
> dirtied=249
>                             ->  Index Scan using gifts_date_added on 
> gifts lts  (cost=0.56..1869136.25 rows=75 width=42) (actual 
> time=190.657..179870.165 rows=6731 loops=1)
>                                   Index Cond: ((date_added > '2018-07-14 
> 11:13:05'::timestamp without time zone) AND (date_added < '2018-08-13 
> 14:14:21'::timestamp without time zone))
>                                   Filter: ((date_gifted >= '2018-08-13 
> 11:13:05'::timestamp without time zone) AND (date_gifted < '2018-08-13 
> 14:14:21'::timestamp without time zone))
>                                   Rows Removed by Filter: 938197
>                                   Buffers: shared hit=1654591 
> read=331897 dirtied=249
>   Planning time: 0.426 ms
>   Execution time: 179893.894 ms
> 
> I don't have a version of this query from prior to this summer, but 
> getting explain plan for older data from older sandboxes show a similar 
> plan.

I don't have an answer, just a question:

Can you provide the actual query and the table schema?

> 
> Sidenote: I am suggesting that an index be added on the date_gifted 
> field as that is far more selective and avoids throwing rows away. 
> However I'm very interested in why every node dealing with the gifts 
> table thinks rows=75 when the actual is much, much higher. And 75 seems 
> like too round of a number to be random?
> 
> -- 
> Don Seiler
> www.seiler.us <http://www.seiler.us>


-- 
Adrian Klaver
adrian.klaver@aklaver.com

On 08/15/2018 01:03 PM, Don Seiler wrote:
> Here's the query, obfuscated manually by me:
> 
> SELECT
>          'Foo' as system_function,
> stores.name <http://stores.name> as store,
>          lt.owner,
>          lt.minute_of_day,
>          lt.records
>          FROM
>          foo.stores
>          LEFT OUTER JOIN
>              (SELECT
>                  lts.store_pkey,
>                  lts.owner,
>                  date_trunc('minute', lts.date_gifted) as minute_of_day,
>                  count(*) as records
>              FROM foo.gifts lts
>              WHERE
>                  lts.date_added  > '2017-07-14 11:13:05'
>              AND lts.date_added  < '2017-08-13 14:14:21'
>              AND lts.date_gifted >= '2017-08-13 11:13:05'
>              AND lts.date_gifted <  '2017-08-13 14:14:21'
>              GROUP BY 1,2,3
>              ORDER BY 1
>              ) lt ON lt.store_pkey = stores.pkey
>          WHERE lt.records IS NOT NULL;
> 
> The foo.gifts table is pretty much the core table of our database. It's 
> big and very active. There is an index on date_added but not yet on 
> date_gifted.
> 
> I'm working to re-write the query while the dev sees if we even need 
> this query anymore.
> 

I agree the issue seems to be in the index/filter of the dates. That 
leads me to another question:

Why in:

WHERE
    lts.date_added  > '2017-07-14 11:13:05'
AND
    lts.date_added  < '2017-08-13 14:14:21'
AND
    lts.date_gifted >= '2017-08-13 11:13:05'
AND
    lts.date_gifted <  '2017-08-13 14:14:21'

is

    lts.date_added  > '2017-07-14 11:13:05'

and
    
    lts.date_gifted >= '2017-08-13 11:13:05'
?

In other words one '>' and the other '>=' ?








-- 
Adrian Klaver
adrian.klaver@aklaver.com

Don Seiler wrote:
> We have a report query that has gone from maybe a few seconds to run to a few minutes to run since mid-July.
> Looking at the output of EXPLAIN ANALYZE, the row count estimates are way off, even though this table was
> just analyzed a day or so ago. What's more bizarre to me is that the row count esimate is *always* 75 for
> every node of the plan, where the actual rows is in the hundreds or thousands. This table is one of the
> busiest tables in our production database (many inserts and updates). It is autovacuumed and autoanalyzed
> a few times per week, although I'm looking to change it to a nightly manual schedule to avoid daytime autovacuums.
> 
>  Hash Join  (cost=1869142.34..1869146.15 rows=75 width=88) (actual time=179877.869..179878.011 rows=759 loops=1)
>    Hash Cond: (stores.pkey = lt.store_pkey)
>    Buffers: shared hit=1654593 read=331897 dirtied=249
>    ->  Seq Scan on stores  (cost=0.00..2.77 rows=77 width=22) (actual time=0.007..0.023 rows=78 loops=1)
>          Buffers: shared hit=2
>    ->  Hash  (cost=1869141.40..1869141.40 rows=75 width=50) (actual time=179877.847..179877.847 rows=759 loops=1)
>          Buckets: 1024  Batches: 1  Memory Usage: 73kB
>          Buffers: shared hit=1654591 read=331897 dirtied=249
>          ->  Subquery Scan on lt  (cost=1869138.59..1869141.40 rows=75 width=50) (actual time=179875.976..179877.697
rows=759loops=1)
 
>                Buffers: shared hit=1654591 read=331897 dirtied=249
>                ->  GroupAggregate  (cost=1869138.59..1869140.65 rows=75 width=50) (actual time=179875.976..179877.606
rows=759loops=1)
 
>                      Group Key: lts.store_pkey, lts.owner, (date_trunc('minute'::text, lts.date_gifted))
>                      Filter: (count(*) IS NOT NULL)
>                      Buffers: shared hit=1654591 read=331897 dirtied=249
>                      ->  Sort  (cost=1869138.59..1869138.78 rows=75 width=42) (actual time=179875.961..179876.470
rows=6731loops=1)
 
>                            Sort Key: lts.store_pkey, lts.entry_source_owner, (date_trunc('minute'::text,
lts.date_gifted))
>                            Sort Method: quicksort  Memory: 757kB
>                            Buffers: shared hit=1654591 read=331897 dirtied=249
>                            ->  Index Scan using gifts_date_added on gifts lts  (cost=0.56..1869136.25 rows=75
width=42)(actual time=190.657..179870.165 rows=6731 loops=1)
 
>                                  Index Cond: ((date_added > '2018-07-14 11:13:05'::timestamp without time zone) AND
(date_added< '2018-08-13 14:14:21'::timestamp without time zone))
 
>                                  Filter: ((date_gifted >= '2018-08-13 11:13:05'::timestamp without time zone) AND
(date_gifted< '2018-08-13 14:14:21'::timestamp without time zone))
 
>                                  Rows Removed by Filter: 938197
>                                  Buffers: shared hit=1654591 read=331897 dirtied=249
>  Planning time: 0.426 ms
>  Execution time: 179893.894 ms
> 
> I don't have a version of this query from prior to this summer, but getting explain plan for older data from
> older sandboxes show a similar plan.
> 
> Sidenote: I am suggesting that an index be added on the date_gifted field as that is far more selective and avoids
> throwing rows away. However I'm very interested in why every node dealing with the gifts table thinks rows=75
> when the actual is much, much higher. And 75 seems like too round of a number to be random?

Yes, I would say that adding an index on "date_gifted" would help.  You may end
up with two bitmap index scans that get combined.
Make sure "work_mem" is big enough to avoid lossy bitmaps (indicated in the plan).

About the misestimate:

You could try running ANALYZE with an increased "default_statistics_target" and see
if that changes the estimate.
If yes, then maybe you should increase statistics for that table or (seing that you are
querying current values) you should collect statistics more often.

Yours,
Laurenz Albe
-- 
Cybertec | https://www.cybertec-postgresql.com