postgres 上的慢速选择不同查询

发布于 2024-11-08 03:53:04 字数 2338 浏览 0 评论 0原文

我在一个基本上收集日志信息的表上经常执行以下两个查询。两者都从大量行中选择不同的值,但其中的不同值少于 10 个。

我分析了页面执行的两个“不同”查询:

marchena=> explain select distinct auditrecor0_.bundle_id as col_0_0_ from audit_records auditrecor0_;
                                          QUERY PLAN                                          
----------------------------------------------------------------------------------------------
 HashAggregate  (cost=1070734.05..1070734.11 rows=6 width=21)
   ->  Seq Scan on audit_records auditrecor0_  (cost=0.00..1023050.24 rows=19073524 width=21)
(2 rows)

marchena=> explain select distinct auditrecor0_.server_name as col_0_0_ from audit_records auditrecor0_;
                                          QUERY PLAN                                          
----------------------------------------------------------------------------------------------
 HashAggregate  (cost=1070735.34..1070735.39 rows=5 width=13)
   ->  Seq Scan on audit_records auditrecor0_  (cost=0.00..1023051.47 rows=19073547 width=13)
(2 rows)

两者都对列进行顺序扫描。但是,如果我关闭enable_seqscan(尽管名称如此,这只会禁用对具有索引的列进行序列扫描)查询将使用索引,但速度会更慢:

marchena=> set enable_seqscan = off;
SET
marchena=> explain select distinct auditrecor0_.bundle_id as col_0_0_ from audit_records auditrecor0_;
                                                       QUERY PLAN                                                       
------------------------------------------------------------------------------------------------------------------------
 Unique  (cost=0.00..19613740.62 rows=6 width=21)
   ->  Index Scan using audit_bundle_idx on audit_records auditrecor0_  (cost=0.00..19566056.69 rows=19073570 width=21)
(2 rows)

marchena=> explain select distinct auditrecor0_.server_name as col_0_0_ from audit_records auditrecor0_;
                                                       QUERY PLAN                                                       
------------------------------------------------------------------------------------------------------------------------
 Unique  (cost=0.00..45851449.96 rows=5 width=13)
   ->  Index Scan using audit_server_idx on audit_records auditrecor0_  (cost=0.00..45803766.04 rows=19073570 width=13)
(2 rows)

bundle_id 和 server_name 列上都有 btree 索引,我应该使用不同的类型吗索引可以快速选择不同的值?

I'm doing the following two queries quite frequently on a table that essentially gathers up logging information. Both select distinct values from a huge number of rows but with less than 10 different values in those.

I've analyzed both "distinct" queries done by the page:

marchena=> explain select distinct auditrecor0_.bundle_id as col_0_0_ from audit_records auditrecor0_;
                                          QUERY PLAN                                          
----------------------------------------------------------------------------------------------
 HashAggregate  (cost=1070734.05..1070734.11 rows=6 width=21)
   ->  Seq Scan on audit_records auditrecor0_  (cost=0.00..1023050.24 rows=19073524 width=21)
(2 rows)

marchena=> explain select distinct auditrecor0_.server_name as col_0_0_ from audit_records auditrecor0_;
                                          QUERY PLAN                                          
----------------------------------------------------------------------------------------------
 HashAggregate  (cost=1070735.34..1070735.39 rows=5 width=13)
   ->  Seq Scan on audit_records auditrecor0_  (cost=0.00..1023051.47 rows=19073547 width=13)
(2 rows)

Both do sequence scans of the columns. However if I turn off enable_seqscan (dispite the name this only disables doing sequence scans on columns with indices) the query uses the index, but is even slower:

marchena=> set enable_seqscan = off;
SET
marchena=> explain select distinct auditrecor0_.bundle_id as col_0_0_ from audit_records auditrecor0_;
                                                       QUERY PLAN                                                       
------------------------------------------------------------------------------------------------------------------------
 Unique  (cost=0.00..19613740.62 rows=6 width=21)
   ->  Index Scan using audit_bundle_idx on audit_records auditrecor0_  (cost=0.00..19566056.69 rows=19073570 width=21)
(2 rows)

marchena=> explain select distinct auditrecor0_.server_name as col_0_0_ from audit_records auditrecor0_;
                                                       QUERY PLAN                                                       
------------------------------------------------------------------------------------------------------------------------
 Unique  (cost=0.00..45851449.96 rows=5 width=13)
   ->  Index Scan using audit_server_idx on audit_records auditrecor0_  (cost=0.00..45803766.04 rows=19073570 width=13)
(2 rows)

Both bundle_id and server_name columns have btree indices on them, should I be using a different type of index to make selecting distinct values fast?

如果你对这篇内容有疑问,欢迎到本站社区发帖提问 参与讨论,获取更多帮助,或者扫码二维码加入 Web 技术交流群。

扫码二维码加入Web技术交流群

发布评论

需要 登录 才能够评论, 你可以免费 注册 一个本站的账号。

评论(4

稚然 2024-11-15 03:53:04
BEGIN; 
CREATE TABLE dist ( x INTEGER NOT NULL ); 
INSERT INTO dist SELECT random()*50 FROM generate_series( 1, 5000000 ); 
COMMIT;
CREATE INDEX dist_x ON dist(x);


VACUUM ANALYZE dist;
EXPLAIN ANALYZE SELECT DISTINCT x FROM dist;

HashAggregate  (cost=84624.00..84624.51 rows=51 width=4) (actual time=1840.141..1840.153 rows=51 loops=1)
   ->  Seq Scan on dist  (cost=0.00..72124.00 rows=5000000 width=4) (actual time=0.003..573.819 rows=5000000 loops=1)
 Total runtime: 1848.060 ms

PG(还)不能使用不同的索引(跳过相同的值),但你可以这样做:

CREATE OR REPLACE FUNCTION distinct_skip_foo()
RETURNS SETOF INTEGER
LANGUAGE plpgsql STABLE 
AS $
DECLARE
    _x  INTEGER;
BEGIN
    _x := min(x) FROM dist;
    WHILE _x IS NOT NULL LOOP
        RETURN NEXT _x;
        _x := min(x) FROM dist WHERE x > _x;
    END LOOP;
END;
$ ;

EXPLAIN ANALYZE SELECT * FROM distinct_skip_foo();
Function Scan on distinct_skip_foo  (cost=0.00..260.00 rows=1000 width=4) (actual time=1.629..1.635 rows=51 loops=1)
 Total runtime: 1.652 ms
BEGIN; 
CREATE TABLE dist ( x INTEGER NOT NULL ); 
INSERT INTO dist SELECT random()*50 FROM generate_series( 1, 5000000 ); 
COMMIT;
CREATE INDEX dist_x ON dist(x);


VACUUM ANALYZE dist;
EXPLAIN ANALYZE SELECT DISTINCT x FROM dist;

HashAggregate  (cost=84624.00..84624.51 rows=51 width=4) (actual time=1840.141..1840.153 rows=51 loops=1)
   ->  Seq Scan on dist  (cost=0.00..72124.00 rows=5000000 width=4) (actual time=0.003..573.819 rows=5000000 loops=1)
 Total runtime: 1848.060 ms

PG can't (yet) use an index for distinct (skipping the identical values) but you can do this :

CREATE OR REPLACE FUNCTION distinct_skip_foo()
RETURNS SETOF INTEGER
LANGUAGE plpgsql STABLE 
AS $
DECLARE
    _x  INTEGER;
BEGIN
    _x := min(x) FROM dist;
    WHILE _x IS NOT NULL LOOP
        RETURN NEXT _x;
        _x := min(x) FROM dist WHERE x > _x;
    END LOOP;
END;
$ ;

EXPLAIN ANALYZE SELECT * FROM distinct_skip_foo();
Function Scan on distinct_skip_foo  (cost=0.00..260.00 rows=1000 width=4) (actual time=1.629..1.635 rows=51 loops=1)
 Total runtime: 1.652 ms
蓦然回首 2024-11-15 03:53:04

您从整个表中选择不同的值,这会自动导致顺序扫描。你有数百万行,所以它一定会很慢。

有一个技巧可以更快地获取不同的值,但它仅在数据具有已知(且相当小)的可能值集时才有效。例如,我认为您的bundle_id引用了某种较小的bundles表。这意味着您可以编写:

select bundles.bundle_id
from bundles
where exists (
      select 1 from audit_records
      where audit_records.bundle_id = bundles.bundle_id
      );

这应该导致对包进行嵌套循环/序列扫描 ->使用bundle_id 上的索引对audit_records 进行索引扫描。

You're selecting distinct values from the whole table, which automatically leads to a seq scan. You've millions rows, so it'll necessarily be slow.

There's a trick to get the distinct values faster, but it only works when the data has a known (and reasonably small) set of possible values. For instance, I take it that your bundle_id references some kind of bundles table which is a smaller. This means you can write:

select bundles.bundle_id
from bundles
where exists (
      select 1 from audit_records
      where audit_records.bundle_id = bundles.bundle_id
      );

This should lead to a nested loop / seq scan on bundles -> index scan on audit_records using the index on bundle_id.

携君以终年 2024-11-15 03:53:04

我对表也有同样的问题> 3 亿条记录和一个包含几个不同值的索引字段。我无法摆脱 seq 扫描,因此我使用此函数来模拟使用索引(如果存在)的不同搜索。如果您的表有许多与记录总数成比例的不同值,则此函数不好。它还必须针对多列不同值进行调整。 警告:此函数对 SQL 注入完全开放,只能在安全环境中使用。

解释分析结果:
使用正常 SELECT DISTINCT 进行查询:总运行时间:598310.705 ms
使用 SELECT Small_distinct(...) 进行查询:总运行时间:1.156 毫秒

CREATE OR REPLACE FUNCTION small_distinct(
   tableName varchar, fieldName varchar, sample anyelement = ''::varchar)
   -- Search a few distinct values in a possibly huge table
   -- Parameters: tableName or query expression, fieldName,
   --             sample: any value to specify result type (defaut is varchar)
   -- Author: T.Husson, 2012-09-17, distribute/use freely
   RETURNS TABLE ( result anyelement ) AS
$BODY$
BEGIN
   EXECUTE 'SELECT '||fieldName||' FROM '||tableName||' ORDER BY '||fieldName
      ||' LIMIT 1'  INTO result;
   WHILE result IS NOT NULL LOOP
      RETURN NEXT;
      EXECUTE 'SELECT '||fieldName||' FROM '||tableName
         ||' WHERE '||fieldName||' > $1 ORDER BY ' || fieldName || ' LIMIT 1'
         INTO result USING result;
   END LOOP;
END;
$BODY$ LANGUAGE plpgsql VOLATILE;

调用示例:

SELECT small_distinct('observations','id_source',1);
SELECT small_distinct('(select * from obs where id_obs > 12345) as temp',
   'date_valid','2000-01-01'::timestamp);
SELECT small_distinct('addresses','state');

I have the same problem with tables > 300 millions records and an indexed field with a few distinct values. I couldn't get rid of the seq scan so I made this function to simulate a distinct search using the index if it exists. If your table has a number of distinct values proportional to the total number of records, this function isn't good. It also has to be adjusted for multi-columns distinct values. Warning: This function is wide open to sql injection and should only be used in a securized environment.

Explain analyze results:
Query with normal SELECT DISTINCT: Total runtime: 598310.705 ms
Query with SELECT small_distinct(...): Total runtime: 1.156 ms

CREATE OR REPLACE FUNCTION small_distinct(
   tableName varchar, fieldName varchar, sample anyelement = ''::varchar)
   -- Search a few distinct values in a possibly huge table
   -- Parameters: tableName or query expression, fieldName,
   --             sample: any value to specify result type (defaut is varchar)
   -- Author: T.Husson, 2012-09-17, distribute/use freely
   RETURNS TABLE ( result anyelement ) AS
$BODY$
BEGIN
   EXECUTE 'SELECT '||fieldName||' FROM '||tableName||' ORDER BY '||fieldName
      ||' LIMIT 1'  INTO result;
   WHILE result IS NOT NULL LOOP
      RETURN NEXT;
      EXECUTE 'SELECT '||fieldName||' FROM '||tableName
         ||' WHERE '||fieldName||' > $1 ORDER BY ' || fieldName || ' LIMIT 1'
         INTO result USING result;
   END LOOP;
END;
$BODY$ LANGUAGE plpgsql VOLATILE;

Call samples:

SELECT small_distinct('observations','id_source',1);
SELECT small_distinct('(select * from obs where id_obs > 12345) as temp',
   'date_valid','2000-01-01'::timestamp);
SELECT small_distinct('addresses','state');
一身骄傲 2024-11-15 03:53:04

在 PostgreSQL 9.3 上,从 Denis 的回答开始:

    select bundles.bundle_id
    from bundles
    where exists (
      select 1 from audit_records
      where audit_records.bundle_id = bundles.bundle_id
      );

只需向子查询添加“限制 1”,我就获得了 60 倍的加速(对于我的用例,有 800 万条记录、复合索引和 10k 组合),从 1800 毫秒开始至 30 毫秒:

    select bundles.bundle_id
    from bundles
    where exists (
      select 1 from audit_records
      where audit_records.bundle_id = bundles.bundle_id limit 1
      );

On PostgreSQL 9.3, starting from the answer from Denis:

    select bundles.bundle_id
    from bundles
    where exists (
      select 1 from audit_records
      where audit_records.bundle_id = bundles.bundle_id
      );

just by adding a 'limit 1' to the subquery, I got a 60x speedup (for my use case, with 8 million records, a composite index and 10k combinations), going from 1800ms to 30ms:

    select bundles.bundle_id
    from bundles
    where exists (
      select 1 from audit_records
      where audit_records.bundle_id = bundles.bundle_id limit 1
      );
~没有更多了~
我们使用 Cookies 和其他技术来定制您的体验包括您的登录状态等。通过阅读我们的 隐私政策 了解更多相关信息。 单击 接受 或继续使用网站,即表示您同意使用 Cookies 和您的相关数据。
原文