Обсуждение: SELECT from a set of values really slow?

I've been benchmarking some very simple databases.  By simple, I mean a table
like this:

    CREATE TABLE bench (
        id SERIAL,
        data TEXT )
    CREATE INDEX bench_data_index ON bench (data)

which is filled with 100k records, where the data values for each record are
distinct (the data for record N is "text_item_N").

I generate a random data value known to be in the table and lookup its id,
with "SELECT id FROM bench WHERE data = 'whatever'".

I run this for three seconds.  PostgreSQL does OK: 4300 selects in 2999
msec.  MySQL does a little better, but not much: 5500 selects in 2999 msec.

In my application, this kind of lookup turns out to be the bottleneck on
performance (which is why I'm benchmarking).  On MySQL, I can speed it up
quite a bit by looking up more than one thing at a time:

    SELECT id FROM bench WHERE data IN ('X', 'Y')

for example, gets about 4050 selects in 3 seconds, so that's 8100 records
looked up, compared to 5300 when they were done one at a time.  It continues
to improve selecting more at a time.  To my surprise, when I tried this
trick with PostgreSQL, it did not speed things up.  In fact, it *massively*
slowed down--it only is getting 13 selects in 3 seconds, searching for two
at a time.

What's going on here?

PostgreSQL 7.4.2 on SuSE 9.1 Linux.  (The version that comes with SuSE).

--
--Tim Smith

hi,

Tim Smith wrote:
>
> What's going on here?
>

#>explain analyze SELECT id FROM bench WHERE data IN ('X', 'Y');
#>set enable_seqscan to off;
#>explain analyze SELECT id FROM bench WHERE data IN ('X', 'Y');

result?

C.

Tim Smith <reply_in_group@mouse-potato.com> writes:
>     SELECT id FROM bench WHERE data IN ('X', 'Y')

> To my surprise, when I tried this trick with PostgreSQL, it did not
> speed things up.  In fact, it *massively* slowed down--it only is
> getting 13 selects in 3 seconds, searching for two at a time.

> What's going on here?

Likely it's switching from index to sequential scan because of a poor
estimate of how many rows will be returned.  Have you ever ANALYZEd
the test table?  Without either ANALYZE stats or a unique index,
the planner will certainly not think that the column is unique.

            regards, tom lane

test=> insert into bench (id,data) select id, 'text_item_'||id::text from
dummy where id<=100000 order by id;
INSERT 0 100001
test=> CREATE INDEX bench_data_index ON bench (data);
CREATE INDEX
test=> explain select * from bench where data = 'test_item_1';
  Index Scan using bench_data_index on bench  (cost=0.00..1478.85 rows=501
width=36)
    Index Cond: (data = 'test_item_1'::text)
(2 lignes)

test=> explain select * from bench where data in ( 'test_item_1',
'test_item_2' );
  Seq Scan on bench  (cost=0.00..2190.01 rows=998 width=36)
    Filter: ((data = 'test_item_1'::text) OR (data = 'test_item_2'::text))
(2 lignes)

test=> vacuum analyze bench;
VACUUM
test=> explain select * from bench where data in ( 'test_item_1',
'test_item_2' );
  Index Scan using bench_data_index, bench_data_index on bench
(cost=0.00..7.91 rows=2 width=22)
    Index Cond: ((data = 'test_item_1'::text) OR (data =
'test_item_2'::text))
(2 lignes)

=> once you analyze, it works...


***********************************************

With 1 item :

test=>explain analyze select * from bench where data in ( 'test_item_1' );
  Total runtime: 0.127 ms

With 11 items :

test=>explain analyze select * from bench where data in ( 'test_item_1',
'test_item_2', 'test_item_55', 'test_item_64', 'test_item_1005',
'test_item_78541', 'test_item_96521', 'test_item_8574', 'test_item_89652',
'test_item_14527', 'test_item_48652' );
  Total runtime: 0.352 ms

***********************************************

With a Join... see on psql-performance



On Sun, 10 Oct 2004 16:00:10 -0400, Tom Lane <tgl@sss.pgh.pa.us> wrote:

> Tim Smith <reply_in_group@mouse-potato.com> writes:
>>     SELECT id FROM bench WHERE data IN ('X', 'Y')
>
>> To my surprise, when I tried this trick with PostgreSQL, it did not
>> speed things up.  In fact, it *massively* slowed down--it only is
>> getting 13 selects in 3 seconds, searching for two at a time.
>
>> What's going on here?
>
> Likely it's switching from index to sequential scan because of a poor
> estimate of how many rows will be returned.  Have you ever ANALYZEd
> the test table?  Without either ANALYZE stats or a unique index,
> the planner will certainly not think that the column is unique.
>
>             regards, tom lane
>
> ---------------------------(end of broadcast)---------------------------
> TIP 5: Have you checked our extensive FAQ?
>
>                http://www.postgresql.org/docs/faqs/FAQ.html
>