PostgreSQL 绑定变量窥探
今天我们要探讨的是 custom执行计划和通用执行计划。这一技术在 Oracle中被称为绑定变量窥视。但 Kingbase中并没有这样的定义,更严格地说,Kingbase叫做custom执行计划和通用执行计划。
什么是custom执行计划,什么是通用执行计划,我们先来看一个例子,我创建了一个100011行的表,其中有两列分别为 id、 name。在name列就2种类型的值,一种值为“aaa”,有整整100000行, 而值为bbb列的仅有11行。这就是我们常说的数据倾斜。在oracle数据库中,配合绑定变量窥视我们常常需要收集倾斜列的直方图。
以下测试基于版本:
KingbaseES V008R006C005B0041 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 4.1.2 20080704 (Red Hat 4.1.2-46), 64-bit
create table a(id numeric,name varchar(40));
insert into a select i, 'aaa' from generate_series (1,100000) i;
insert into a select i, 'bbb' from generate_series (100001,100011) i;
create index idx_a1 on a(name);
analyze a;
下一步是使用 prepare语句。利用该方法可以避免对语句反复解析。这个功能类似oracle 的绑定变量,(一次硬解析后在library cache产生的执行计划可为以后sql通用。避免多次硬解析,这样找到相同的执行计划planhash value叫做软解析。当然还有软软解析,这里略过。)
test=# prepare test_stmt as select * from a where name = $1;
PREPARE
select * from pg_prepared_statements;
我们执行如下语句,连续6次都查询name为'aaa'的数据。注意是6次。
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
---------------------------------------------------------------------------------------------------------
Seq Scan on a (cost=0.00..1794.14 rows=99994 width=10) (actual time=0.009..25.862 rows=100000 loops=1)
Filter: ((name)::text = 'aaa'::text)
Rows Removed by Filter: 11
Planning Time: 0.217 ms
Execution Time: 34.710 ms
(5 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
---------------------------------------------------------------------------------------------------------
Seq Scan on a (cost=0.00..1794.14 rows=99994 width=10) (actual time=0.009..16.401 rows=100000 loops=1)
Filter: ((name)::text = 'aaa'::text)
Rows Removed by Filter: 11
Planning Time: 0.073 ms
Execution Time: 23.340 ms
(5 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
---------------------------------------------------------------------------------------------------------
Seq Scan on a (cost=0.00..1794.14 rows=99994 width=10) (actual time=0.009..30.001 rows=100000 loops=1)
Filter: ((name)::text = 'aaa'::text)
Rows Removed by Filter: 11
Planning Time: 0.093 ms
Execution Time: 39.383 ms
(5 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
---------------------------------------------------------------------------------------------------------
Seq Scan on a (cost=0.00..1794.14 rows=99994 width=10) (actual time=0.009..23.365 rows=100000 loops=1)
Filter: ((name)::text = 'aaa'::text)
Rows Removed by Filter: 11
Planning Time: 0.073 ms
Execution Time: 32.397 ms
(5 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
---------------------------------------------------------------------------------------------------------
Seq Scan on a (cost=0.00..1794.14 rows=99994 width=10) (actual time=0.009..19.287 rows=100000 loops=1)
Filter: ((name)::text = 'aaa'::text)
Rows Removed by Filter: 11
Planning Time: 0.099 ms
Execution Time: 27.462 ms
(5 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..1710.52 rows=50006 width=10) (actual time=0.082..35.540 rows=100000 loops=1)
Index Cond: ((name)::text = $1)
Planning Time: 0.114 ms
Execution Time: 45.546 ms
(4 rows)
由于 aaa占用了该表的大部分数据,因此优化器选择使用全表扫描,这是优化器的算法决定的,这也存在合理性。在第六次的时候,请注意 Filter部分,(name)::text = 'aaa'::text变为 text=$1。此时优化器将生成通用执行计划,并使用绑定变量。那么之前的5次则被称为 custom执行计划。为什么第六次才生成通用执行计划?我们可以在 PostgreSQL的 plancache. c源码中找到说明:
The logic for choosing generic or custom plans is in choose_custom_plan
在choose_custom_plan函数里我们可以看到/* Generate costom plans until we have done at least 5 (arbitrary)*/ if (planaource->num_custom_plans < 5) return true;
请注意,这里的限定值小于5次,返回 true,选择 custom执行计划,而大于5次之后,则选择通用执行计划。因此,5次之后执行计划就会固定。为什么第六次使用通用执行计划,执行计划却改为索引扫描的方式?实际上这和一个参数有关plan_cache_mode。目前查看参数值时auto。
test=# show plan_cache_mode;
plan_cache_mode
-----------------
auto
(1 row)
在参数是auto的前提下,不管我执行aaa或bbb的列值,执行计划都是一样,执行计划固定了。如果每次不管变量值怎么变化,都选择索引扫描方式,显然这不是我们想要的。因为数据倾斜,如果执行计划不变,那么是不明智的,会出现低效解析行为。
如下,使用通用执行计划后,我们关注不管索引扫描还是全表扫描,预估cost值是50006,有意思的是这个值是实际rows的一半。从第六次执行计划开始,这个cost就没再变过,显然这是不合理的。当然有可能优化器认为这种算法对于不同的扫描方式对应的Execution Time差的不是很多,所以固定执行计划为通用执行计划。
还有一个关键是使用通用执行计划后Planning Time很小,这是否说明了”软解析的功能呢!“生成执行计划时间大大减少。
test=# explain (analyze) execute test_stmt ('bbb');
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..1710.52 rows=50006 width=10) (actual time=0.021..0.024 rows=11 loops=1)
Index Cond: ((name)::text = $1)
Planning Time: 0.015 ms
Execution Time: 0.048 ms
(4 rows)
test=# explain (analyze) execute test_stmt ('bbb');
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..1710.52 rows=50006 width=10) (actual time=0.020..0.022 rows=11 loops=1)
Index Cond: ((name)::text = $1)
Planning Time: 0.014 ms
Execution Time: 0.041 ms
(4 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..1710.52 rows=50006 width=10) (actual time=0.032..23.592 rows=100000 loops=1)
Index Cond: ((name)::text = $1)
Planning Time: 0.013 ms
Execution Time: 31.333 ms
(4 rows)
设置 plan_cache_mode=force_custom_plan
继续测试另外一种情况,将plan_cache_mode设置为force_custom_plan。可以看到执行计划会根据绑定变量的值的分布进行变化,这种情况执行计划是合理的。但是代价是每次执行都要重新解析语句,我们知道在oracle里这叫硬解析,都听说过一句话,硬解析是万恶之源!对应的在Kingbase里数据倾斜,谓词条件经常变化,最好使用custom执行计划。
set plan_cache_mode=force_custom_plan;
test=# explain (analyze) execute test_stmt ('bbb');
QUERY PLAN
-------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..8.65 rows=13 width=10) (actual time=0.020..0.022 rows=11 loops=1)
Index Cond: ((name)::text = 'bbb'::text)
Planning Time: 0.079 ms
Execution Time: 0.036 ms
(4 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
---------------------------------------------------------------------------------------------------------
Seq Scan on a (cost=0.00..1794.14 rows=99998 width=10) (actual time=0.010..16.897 rows=100000 loops=1)
Filter: ((name)::text = 'aaa'::text)
Rows Removed by Filter: 11
Planning Time: 0.077 ms
Execution Time: 24.020 ms
(5 rows)
设置 plan_cache_mode=force_generic_plan
可以看到,这种情况下,执行计划就被固定了。和最开始执行到第六次的执行计划一样,不管条件怎么变化,优化器都采用了通用执行计划。
test=# set plan_cache_mode =force_generic_plan ;
test=# explain (analyze) execute test_stmt ('bbb');
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..1710.52 rows=50006 width=10) (actual time=0.022..0.024 rows=11 loops=1)
Index Cond: ((name)::text = $1)
Planning Time: 0.016 ms
Execution Time: 0.044 ms
(4 rows)
test=# explain (analyze) execute test_stmt ('bbb');
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..1710.52 rows=50006 width=10) (actual time=0.032..0.035 rows=11 loops=1)
Index Cond: ((name)::text = $1)
Planning Time: 0.015 ms
Execution Time: 0.055 ms
(4 rows)
test=# explain (analyze) execute test_stmt ('aaa');
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------
Index Scan using idx_a1 on a (cost=0.42..1710.52 rows=50006 width=10) (actual time=0.037..23.191 rows=100000 loops=1)
Index Cond: ((name)::text = $1)
Planning Time: 0.016 ms
Execution Time: 30.997 ms
(4 rows)
关闭prepare语句
deallocate all;
结论:
如果在Kingbase中使用prepare语句(类似绑定变量功能),
对于数据分布均匀,且参数经常改变的情况适合使用这个功能。
建议对于数据倾斜的情况,将plan_cache_mode设置为force_custom_plan。或者不用这个功能。
当然在实现任何功能前还是建议进行充分测试。