Re: [GENERAL] Significant discrepancy in index cost estimation

After setting seq_page_cost to 3 the execution plan became good, without SeqScan, but it seems strange to set seq_page_cost almost equal to random_page_cost, therefore i've set seq_page_cost back to defaults, increased the statistics for "sub_id" in "mba_test.subscr_param" to 1000. That gave me the value of estimated number of rows for subscr_param_pkey closer to the real value (est.64, real 30) which affected the execution plan in the right way.

>Thursday, June 29, 2017 7:21 PM +03:00 from Mikhail <bemewe(at)mail(dot)ru>:
>
>
>Hi guys,
>I'm loss. I'm running:
>=# select version();
>                                                 version
>----------------------------------------------------------------------------------------------------------
>PostgreSQL 9.6.2 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 4.8.5 20150623 (Red Hat 4.8.5-11), 64-bit
> 
>=# show work_mem;
>work_mem
>----------
>27MB
> 
>=# show shared_buffers;
>shared_buffers
>----------------
>3520MB
> 
> 
>Having the tables:
>CREATE TABLE mba_test.subscr_param(
>    sub_id integer NOT NULL,
>    res_id integer NOT NULL,
>    rel_status integer,
>    status integer,
>   and so on.. total 35 columns
>   
>CREATE TABLE mba_test.subscr (
>    sub_id integer NOT NULL,
>    sub_name character varying(80),
>    status integer NOT NULL,
>    acc_id integer,
>   and so on.. total 59 columns
>  
>alter table mba_test.subscr_param add primary key (sub_id, res_id);
>alter table mba_test.subscr add primary key (sub_id);
>create index idx_subscr_acc_id on mba_test.subscr(acc_id);
> 
>Tables and indexes has the following sizes / statistics:
>=# select relname, relpages, reltuples, pg_size_pretty(pg_relation_size(oid))
>     from pg_class
>    where relname in ('subscr_param', 'subscr', 'idx_subscr_acc_id', 'subscr_pkey', 'subscr_param_pkey');
>       relname         | relpages |  reltuples  | pg_size_pretty
>-----------------------+----------+-------------+----------------
>subscr                |    24086 |      825305 | 188 MB
>subscr_param_pkey     |   115968 | 4.22936e+07 | 906 MB
>subscr_param          |  1446158 | 4.22936e+07 | 11 GB
>subscr_pkey           |     2265 |      825305 | 18 MB
>idx_subscr_acc_id     |     2265 |      825305 | 18 MB
> 
>And "subscr" data distribution is:
>=# select acc_id, count(*) from mba_test.subscr group by acc_id order by count(*) desc limit 5;
>  acc_id | count
>---------+-------
>1089212 |  5232
>1000154 |  2884
>1587496 |  1896
>1409682 |  1746
>1249568 |  1149
> 
>=# select count(*) from mba_test.subscr;
>count
>--------
>825305
> 
>=# select count(*) from mba_test.subscr_param;
>  count
>----------
>42293572
> 
>Now, i take the second acc_id (1000154) and run the query below twice (to have cached everything i can). The second execution gives the following:
> 
>=# explain (analyze, buffers) SELECT "SP".res_id, "SP".sub_id
>  FROM mba_test.subscr_param "SP"
>  JOIN mba_test.subscr "S" ON "SP".sub_id = "S".sub_id
>WHERE "S".acc_id = 1000154;
>                                                                     QUERY PLAN
>-----------------------------------------------------------------------------------------------------------------------------------------------------
>Hash Join  (cost=7841.72..2036917.93 rows=138159 width=8) (actual time=39.501..10086.843 rows=86933 loops=1)
>   Hash Cond: ("SP".sub_id = "S".sub_id)
>   Buffers: shared hit=178674 read=1269448
>   ->  Seq Scan on subscr_param "SP"  (cost=0.00..1869093.72 rows=42293572 width=8) (actual time=0.024..6294.100 rows=42293572 loops=1)
>         Buffers: shared hit=176710 read=1269448
>   ->  Hash  (cost=7808.02..7808.02 rows=2696 width=4) (actual time=3.161..3.161 rows=2884 loops=1)
>         Buckets: 4096  Batches: 1  Memory Usage: 134kB
>         Buffers: shared hit=1964
>         ->  Bitmap Heap Scan on subscr "S"  (cost=53.32..7808.02 rows=2696 width=4) (actual time=0.471..2.802 rows=2884 loops=1)
>               Recheck Cond: (acc_id = 1000154)
>               Heap Blocks: exact=1953
>               Buffers: shared hit=1964
>               ->  Bitmap Index Scan on idx_subscr_acc_id  (cost=0.00..52.64 rows=2696 width=0) (actual time=0.273..0.273 rows=2884 loops=1)
>                     Index Cond: (acc_id = 1000154)
>                     Buffers: shared hit=11
>Planning time: 0.155 ms
>Execution time: 10091.265 ms
> 
>Seems strange to decide to seq scan the table with 42 mln rows and size 11GB when having the index (actually, primary key) containing both columns (sub_id and res_id) which is less than 1GB.
> 
>Now, i've explicitly turned the sec scans off and got perfect execution plan:
> 
>=# set enable_seqscan = off;
>=# explain (analyze, buffers) SELECT "SP".res_id, "SP".sub_id
>  FROM mba_test.subscr_param "SP"
>  JOIN mba_test.subscr "S" ON "SP".sub_id = "S".sub_id
>WHERE "S".acc_id = 1000154;
>                                                                           QUERY PLAN
>-----------------------------------------------------------------------------------------------------------------------------------------------------------------
>Nested Loop  (cost=53.88..4954693.91 rows=138159 width=8) (actual time=0.471..62.315 rows=86933 loops=1)
>   Buffers: shared hit=50837
>   ->  Bitmap Heap Scan on subscr "S"  (cost=53.32..7808.02 rows=2696 width=4) (actual time=0.459..3.250 rows=2884 loops=1)
>         Recheck Cond: (acc_id = 1000154)
>         Heap Blocks: exact=1953
>         Buffers: shared hit=1964
>         ->  Bitmap Index Scan on idx_subscr_acc_id  (cost=0.00..52.64 rows=2696 width=0) (actual time=0.258..0.258 rows=2884 loops=1)
>               Index Cond: (acc_id = 1000154)
>               Buffers: shared hit=11
>   ->  Index Only Scan using subscr_param_pkey on subscr_param "SP"  (cost=0.56..1825.67 rows=923 width=8) (actual time=0.004..0.017 rows=30 loops=2884)
>         Index Cond: (sub_id = "S".sub_id)
>         Heap Fetches: 86933
>         Buffers: shared hit=48873
>Planning time: 0.169 ms
>Execution time: 66.644 ms
> 
>67 milliseconds vs 10 seconds! While the cost is two times bigger (4954693 vs 2036917).
>My thoughts are: taking into account that the estimated number of rows to fetch on bitmap heap scan is approx. right (est. 2696 - real 2884), the problem is with index scan cost estimation, which results to 4920200 (2696*1825). And the miss
in cost estimation is because of the estimation of the number of rows to return (est. 923 - real 30).
> 
>And my question:
>  1. am i right in my hypothesis?
>  2. is there a way to affect the cost evaluation in my case to help postgresql choose the right execution plan?
>  3. is there a way to fix this query and not to break the execution of other queries?
> 
>ps: running "analyze" on both tables doesn't affect the result
>pps: all "cost" parameters are in their default values: cpu_index_tuple_cost, cpu_operator_cost, cpu_tuple_cost, random_page_cost, seq_page_cost.
> 
>Thanks, Mikhail
>----------------------------------------------------------------------