postgresql – 在timestamp列上按查询排序非常慢

尽管有索引,并且id为主键,下面的查询大约需要15秒才能返回数据.

select id from my_table order by insert_date offset 0 limit 1

解释分析如下

"Limit  (cost=1766417.72..1766417.72 rows=1 width=12) (actual time=32479.440..32479.441 rows=1 loops=1)"
"  ->  Sort  (cost=1766417.72..1797117.34 rows=12279848 width=12) (actual time=32479.437..32479.437 rows=1 loops=1)"
"        Sort Key: insert_date"
"        Sort Method: top-N heapsort  Memory: 25kB"
"        ->  Seq Scan on my_table  (cost=0.00..1705018.48 rows=12279848 width=12) (actual time=0.006..21338.401 rows=12108916 loops=1)"
"Total runtime: 32479.476 ms"

我的表几乎没有其他专栏.但insert_date的类型是

insert_date timestamp without time zone NOT NULL DEFAULT Now(),

我在该特定日期列上有一个索引

CREATE INDEX my_table_insert_date_indx
  ON my_table
  USING btree
  (insert_date)
TABLESPACE somexyz_idx_ts;

postgresql.conf文件中的值很少：

shared_buffers = more than 1GB    ## just for an example
temp_buffers = more than 1GB
work_mem = more than 1GB
maintenance_work_mem = more than 1GB
dynamic_shared_memory_type = posix
default_statistics_target = 10000
autovacuum = on
random_page_cost = 2.0
cpu_index_tuple_cost = 0.0005

我现在正在使用postgres 9.3.

更新::

我刚刚运行了以下查询：

select insert_date,count(*) from my_table group by insert_date

而结果中的前几位是：

"2015-04-02 00:00:00";3718104
"2015-04-03 00:00:00";6410253
"2015-04-04 00:00:00";538247
"2015-04-05 00:00:00";1228877
"2015-04-06 00:00:00";131248

我在那张桌子上有大约1200万条记录.而上述数字几乎接近总数.

不确定,但是在一个有大量重复值的列上创建索引可能是一个问题吗？如果确实如此,那么我们有什么办法吗？

解决方法

使用Postgresql 9.3和9.4,我的机器上的查询运行速度提高了大约160000倍.我的机器没什么特别的.

-- From Postgresql 9.4; 9.3 is similar.
show shared_buffers; -- 128MB
show temp_buffers; -- 8MB
show work_mem; -- 4MB
show maintenance_work_mem; -- 64MB
show dynamic_shared_memory_type; -- posix
show default_statistics_target; -- 100
show autovacuum; -- on
show random_page_cost; -- 4
show cpu_index_tuple_cost; -- 0.005

制备

让我们建一张桌子. (你应该在你的问题中这样做.)

create table my_table (
  id serial primary key,insert_date timestamp not null
);

-- Round numbers of rows.
insert into my_table(insert_date)
select timestamp '2015-04-02 00:00:00'
from generate_series(1,3000000) n;

insert into my_table(insert_date)
select timestamp '2015-04-03 00:00:00'
from generate_series(1,6000000) n;

insert into my_table(insert_date)
select timestamp '2015-04-04 00:00:00'
from generate_series(1,500000) n;

insert into my_table(insert_date)
select timestamp '2015-04-05 00:00:00'
from generate_series(1,1200000) n;

insert into my_table(insert_date)
select timestamp '2015-04-06 00:00:00'
from generate_series(1,131000) n;

创建索引并更新统计信息.

create index on my_table (insert_date);
analyze my_table;

Postgresql 9.4

现在,我们从您的第一个查询中获得了什么样的执行计划？

explain analyze 
select id from my_table order by insert_date offset 0 limit 1;

"Limit  (cost=0.43..0.48 rows=1 width=12) (actual time=0.014..0.014 rows=1 loops=1)"
"  ->  Index Scan using my_table_insert_date_idx on my_table  (cost=0.43..540656.27 rows=11200977 width=12) (actual time=0.012..0.012 rows=1 loops=1)"
"Planning time: 0.195 ms"
"Execution time: 0.032 ms"

Postgresql 9.3

explain analyze 
select id from my_table order by insert_date offset 0 limit 1;

"Limit  (cost=0.43..0.47 rows=1 width=12) (actual time=0.058..0.059 rows=1 loops=1)"
"  ->  Index Scan using my_table_insert_date_idx on my_table  (cost=0.43..339814.36 rows=10830995 width=12) (actual time=0.057..0.057 rows=1 loops=1)"
"Total runtime: 0.098 ms"

您的查询

select id from my_table order by insert_date offset 0 limit 1;

是不确定的.根据ORDER BY子句,有300万行具有最低的insert_date(首先出现的日期).你挑选了300万中的一个. Postgresql不保证你每次都会得到相同的id.

如果你不关心它返回的300万个ID中的哪一个,你可以用不同的方式表达查询.但我不认为表达不同会给你160k倍的加速.

您可以针对特定查询更改您包含的某些设置.所以你可以做这样的事情.

-- Don't commit or rollback . . . 
begin transaction;
set local work_mem = '8 MB';

explain analyze 
select id from my_table order by insert_date offset 0 limit 1;
-- displays the result.

手动提交或回滚.

commit;

您的work_mem设置返回到服务器启动时设置的值.

show work_mem; -- 4MB