In子查询的原理
1. in原理
此调研的背景是同样的select结果为什么使用子查询会比不使用子查询慢。我们使用的数据库还是mysql官方的employees。进行的实验操作为下面两个语句:
方法一:explain select sql_no_cache
t.emp_no,t.title,t.from_date,t.to_datefrom titles t straight_join
employees e on (e.emp_no=t.emp_no) straight_joinsalaries s on
(s.emp_no=e.emp_no) where e.gender='F' and s.salary=90930;
图1直接使用join
方法二:explain select sql_no_cache * from titles t where t.emp_no
in (select s.emp_no from salaries s, employees e where
s.emp_no=e.emp_no ande.gender='F' and s.salary=90930);
图2使用in的子查询
在下面的讨论中我们直接使用直接join和in的称呼来代表两种不同的情况。(注:在我们的实验中第一种情况use
3.5s;第二种情况use 5.7s)
首先我们来解释一下图2的dependent
subquery是什么意思:手册上的解释是,子查询中的第一个select,取决于外面的查询。就这么一句话,其实它表达的意思是(以我们图2的表来说明)子查询(e
join
t)的第一个表(e)的查询方式依赖于外部(t表)的查询。换句话说就是e表的检索方式依赖于t表的数据,如这里t表得到的记录t.emp_no(where
t.emp_no
in)刚好可以被e表作为eq_ref方式来获得它的相应的记录;换种写法如果此时t表扫描第一条记录得到的t.emp_no为10001的话,那么后面子查询的语句就类似于这样的语句:
select s.emp_no from salaries s, employeese where
s.emp_no=e.emp_no and e.gender='F' and s.salary=90930
ands.emp_no=10001。此时这个语句就会被优化拿来优化,变成了上面的子查询的执行计划。
通过这个解释我们可以知道:对于上面的两种方式,它们使用的索引及读取数据的过程及方法是一样的,全表扫描t表,将t的每条记录传递给e表,e表通过eq_ref索引方式来获得记录判断自身的条件,然后再传递给s给,s表使用ref方式来获得记录,再判断自身的条件是否也得到满足,如果也满足的话,则找到一个满足此查询的语句。那么为什么这两种情况会有性能上的差距了?
首先我们通过bt上看一下,两的具体执行流程,看它们的区别在哪里?
#0 evaluate_join_record (join=0x6fe0f10, join_tab=0x6fe2b28, error=0)
at sql_select.cc:11414
#1 0x00000000005e41e8 in sub_select
(join=0x6fe0f10, join_tab=0x6fe2b28,
end_of_records=)
at sql_select.cc:11384 【s表】
#2 0x00000000005e3f5a in evaluate_join_record (join=0x6fe0f10,
join_tab=0x6fe28d0, error=
out>) at sql_select.cc:11511
#3 0x00000000005e41e8 in sub_select
(join=0x6fe0f10, join_tab=0x6fe28d0,
end_of_records=)
at
sql_select.cc:11384 【e表】
#4 0x00000000005e3f5a in evaluate_join_record (join=0x6fe0f10,
join_tab=0x6fe2678, error=
out>) at sql_select.cc:11511
#5 0x00000000005e4215 in sub_select
(join=0x6fe0f10, join_tab=0x6fe2678,
end_of_records=)
at
sql_select.cc:11391 【t表】
#6 0x0000000000601d30 in do_select
(join=0x6fe0f10, fields=0x6f819a0, table=0x0, procedure=0x0) at
sql_select.cc:11140
#7 0x000000000060a479 in JOIN::exec
(this=0x6fe0f10) at sql_select.cc:2314
#8 0x000000000060ae0f in mysql_select
(thd=0x6f7f980, rref_pointer_array=0x6f81a68, tables=0x6fd5198,
wild_num=0, fields=@0x6f819a0,
conds=0x6fdd218, og_num=0, order=0x0, group=0x0, having=0x0,
proc_param=0x0, select_options=2147764736, result=0x6fdd398,
unit=0x6f81470,
select_lex=0x6f81898) at sql_select.cc:2509
#9 0x000000000060b481 in handle_select
(thd=0x6f7f980, lex=0x6f813d0, result=0x6fdd398,
setup_tables_done_option=0) at sql_select.cc:269
#10 0x000000000054c71a in execute_sqlcom_select (thd=0x6f7f980,
all_tables=0x6fd5198) at sql_parse.cc:5075
#11 0x000000000055538c in mysql_execute_command (thd=0x6f7f980) at
sql_parse.cc:2271
#12 0x000000000055ebd3 in mysql_parse (thd=0x6f7f980,
表1 join方式的bt
通过该bt我们也可以清楚的看到三层nest-loop的过程;注:通过在每一层的sub_select处查看join_tab->table->alias变量我们可以此时具体操作的表。
#0 evaluate_join_record (join=0x6fe11d8, join_tab=0x6fe5148, error=0)
at sql_select.cc:11414
#1 0x00000000005e41e8 in sub_select
(join=0x6fe11d8, join_tab=0x6fe5148,
end_of_records=)
at
sql_select.cc:11384 【s表】
#2 0x00000000005e3f5a in evaluate_join_record (join=0x6fe11d8,
join_tab=0x6fe4ef0, error=
out>) at sql_select.cc:11511
#3 0x00000000005e41e8 in sub_select
(join=0x6fe11d8, join_tab=0x6fe4ef0,
end_of_records=)
at
sql_select.cc:11384 【e表】
#4 0x0000000000601d30 in do_select
(join=0x6fe11d8, fields=0x6fd5300, table=0x0, procedure=0x0) at
sql_select.cc:11140
#5 0x000000000060a479 in JOIN::exec
(this=0x6fe11d8) at sql_select.cc:2314
#6 0x00000000004d5102
in subselect_single_select_engine::exec (this=0x6fdcda0)
at item_subselect.cc:1987
#7 0x00000000004d26b1
in Item_subselect::exec (this=0x6fdccb0)
at item_subselect.cc:280
#8 0x00000000004d1aaa in
Item_in_subselect::val_bool (this=0x6fe11d8) at
item_subselect.cc:880
#9 0x0000000000438821 in Item::val_bool_result
(this=0x6fe11d8) at item.h:745
#10 0x0000000000476941 in Item_in_optimizer::val_int
(this=0x6fe2a58) at item_cmpfunc.cc:1833
#11
0x00000000005e3d9a in evaluate_join_record (join=0x6fdce98,
join_tab=0x6fe3538, error=
out>) at sql_select.cc:11434
#12 0x00000000005e4215 in sub_select (join=0x6fdce98,
join_tab=0x6fe3538, end_of_records=
out>) at
sql_select.cc:11391 【t表】
#13 0x0000000000601d30 in do_select (join=0x6fdce98,
fields=0x6f819a0, table=0x0, procedure=0x0) at
sql_select.cc:11140
#14 0x000000000060a479 in JOIN::exec (this=0x6fdce98) at
sql_select.cc:2314
#15 0x000000000060ae0f in mysql_select (thd=0x6f7f980,
rref_pointer_array=0x6f81a68, tables=0x6fd4d90, wild_num=1,
fields=@0x6f819a0,
conds=0x6fdccb0, og_num=0, order=0x0, group=0x0, having=0x0,
proc_param=0x0, select_options=2147764736, result=0x6fdce78,
unit=0x6f81470,
select_lex=0x6f81898) at sql_select.cc:2509
#16 0x000000000060b481 in handle_select (thd=0x6f7f980,
lex=0x6f813d0, result=0x6fdce78, setup_tables_done_option=0) at
sql_select.cc:269
#17 0x000000000054c71a in execute_sqlcom_select (thd=0x6f7f980,
all_tables=0x6fd4d90) at sql_parse.cc:5075
#18 0x000000000055538c in mysql_execute_command (thd=0x6f7f980) at
sql_parse.cc:2271
#19 0x000000000055ebd3 in mysql_parse (thd=0x6f7f980,
表2 in方式的bt
通过这两个表我们可以发现在表t与表[e,s]之前插入了一些其它的函数,并且这个插入的时机是在t表执行evaluate_join_record函数时调用select_cond_result=
test(select_cond->val_int());判断它所拥有的条件是否满足时进入的。对于表1直接join的情况该过程是没有被执行的因为它没有自身的where
cond。对于表2
in的方式,这个条件可能是由于在前面执行optimize时,确定外部查询的执行计划时确定的,这里我们不再去确认。正常的情况这个test如果有条件的话那么应该执行相应的条件判断如对于e表它最终调用的判断函数为int
Arg_comparator::compare_int_signed();而在这里对于有子查询的它调用的方法是:Item_subselect::exec,而它最终又调用各自的engine->exec(),如这里它调用的是这个subselect_single_select_engine::exec方法,如果是第一次调用该函数的话那么就先执行一次join->optimize(),即对子查询(内部查询)进行优化,而在mysql_select时调用的join->optimize()只是对外部查询进行优化,它并不包括内部查询的优化(执行计划等,另外对于直接join的话没有所谓的内外部查询那么它的整个执行计划就是mysql_select完成),然后执行join->reinit(),最后再执行JOIN::exec;也就是说对于in这种情况t进行全表扫描,那么它总共有443310,那么这几个函数就要被调用443310多次(join->optimize()除外,它在第一次调用子查询确认执行计划之后就不再调用)。
我们可以通过下面的图来反应这两种过程:
图3 join方式的执行过程
图4 in方式的执行过程
上面就是in子查询的实现过程。下面我们将讨论为什么in方式与join方式性能差的原因?
2. In比join慢的原因
首先我们通过oprofile来测试一下,两种情况各自的性能损耗在哪里?
Join
In
samples % symbol
name
444 11.2065 buf_calc_page_new_checksum
337 8.5058 rec_get_offsets_func
326 8.2282 pthread_mutex_unlock
245 6.1837 pthread_mutex_lock
240 6.0575 cmp_dtuple_rec_with_match
226 5.7042 row_search_for_mysql
218 5.5023 row_sel_store_mysql_rec
104 2.6249 code_state
103 2.5997 ha_insert_for_fold
97 2.4483 page_cur_search_with_match
83 2.0949 pthread_mutex_trylock
79 1.9939 pthread_getspecific
77 1.9435 memcpy
76 1.9182 _db_enter_
67 1.6911 btr_search_guess_on_hash
63 1.5901 evaluate_join_record(JOIN*,
st_join_table*,
int)
62 1.5649 safe_mutex_lock
52 1.3125 DoTrace
51 1.2872 Arg_comparator::compare_int_signed()
51 1.2872 _db_return_
49 1.2367 btr_search_build_page_hash_index
46 1.1610 btr_cur_search_to_nth_level
46 1.1610 ha_innobase::general_fetch(unsigned char*,
unsigned int, unsigned int)
37 0.9339 safe_mutex_unlock
33 0.8329 ha_innobase::unlock_row()
32 0.8077 ha_remove_all_nodes_to_page
30 0.7572 btr_search_drop_page_hash_index
28 0.7067 _db_doprnt_
27 0.6815 _db_pargs_
26 0.6562 join_read_key(st_join_table*)
26 0.6562 lock_clust_rec_cons_read_sees
24 0.6058 cp_buffer_from_ref(THD*, st_table*,
st_table_ref*)
24 0.6058 row_get_rec_sys_field
19 0.4796 Field_long::val_int()
19 0.4796 ha_delete_hash_node
19 0.4796 ha_innobase::index_read(unsigned char*,
unsigned char const*, unsigned int, ha_rkey_function)
19 0.4796 mtr_memo_slot_release
19 0.4796 row_sel_convert_mysql_key_to_innobase
18 0.4543 Item_field::val_int()
samples % symbol
name
432 7.4922 pthread_mutex_unlock
423 7.3361 buf_calc_page_new_checksum
373 6.4690 rec_get_offsets_func
283 4.9081 row_search_for_mysql
256 4.4398 pthread_mutex_lock
242 4.1970 _db_enter_
242 4.1970 cmp_dtuple_rec_with_match
206 3.5727 row_sel_store_mysql_rec
190 3.2952 code_state
181 3.1391 _db_return_
151 2.6188 pthread_getspecific
146 2.5321 DoTrace
124 2.1505 build_template(row_prebuilt_struct*,
THD*, st_table*, unsigned
int)
108 1.8730 page_cur_search_with_match
99 1.7170 ha_insert_for_fold
87 1.5088 btr_search_guess_on_hash
83 1.4395 evaluate_join_record(JOIN*,
st_join_table*,
int)
81 1.4048 pthread_mutex_trylock
80 1.3874 safe_mutex_lock
76 1.3181 memcpy
57 0.9886 JOIN::exec()
56 0.9712 __errno_location
54 0.9365 _db_doprnt_
50 0.8672 btr_search_build_page_hash_index
50 0.8672 ha_innobase::general_fetch(unsigned char*,
unsigned int, unsigned int)
47 0.8151 dict_index_copy_types
46 0.7978 btr_cur_search_to_nth_level
41 0.7111 _my_thread_var
39 0.6764 Field_long::val_int()
38 0.6590 Arg_comparator::compare_int_signed()
36 0.6243 safe_mutex_unlock
34 0.5897 cp_buffer_from_ref(THD*, st_table*,
st_table_ref*)
33 0.5723 ha_innobase::index_read(unsigned char*,
unsigned char const*, unsigned int, ha_rkey_function)
32 0.5550 _db_pargs_
31 0.5376 JOIN::cleanup(bool)
31 0.5376 ha_innobase::unlock_row()
29 0.5029 join_read_key(st_join_table*)
29 0.5029 sub_select(JOIN*, st_join_table*,
bool)
通过对比我们发现在join出现的在in中也有,但是在in中占的采样比例比较高的几个在join中并没有(这里的没有并不是指join里没有调用它们,只是可能它们调用的次数太少导致在oprofile采样时没有获得它们的数据)。通过gdb我们发现两种方式都调用build_template,所以为了进一步查看它们的性能损耗,我们通过gcov比较两者的覆盖情况。通过它们的输出我们知道两者调用build_template的次数分别是join
VS in:174
620786;而这个函数的内部有一个for循环用于拷贝,判断哪些字段是需要被保留的。对于in的方式这个for内部执行了3370262次,而join只执行了1036次。我们再来看一下这个函数的bt:
#0 build_template (prebuilt=0x2aaaabbc40b8,
thd=0x6f7f980, table=0x6fd9780, templ_type=1) at
handler/ha_innodb.cc:3564
#1 0x000000000083953b in
ha_innobase::change_active_index (this=0x6fdad30, keynr=0) at
handler/ha_innodb.cc:4834
#2 0x0000000000839ef7 in ha_innobase::index_init
(this=0x6fdad30, keynr=0, sorted=
out>) at handler/ha_innodb.cc:4508
#3 0x00000000006005f8 in join_read_key
(tab=0x6fe4ef0) at handler.h:1180
#4 0x00000000005e41d3 in sub_select
(join=0x6fe11d8, join_tab=0x6f7f980, end_of_records=128) at
sql_select.cc:11383
#5 0x0000000000601d30 in do_select
(join=0x6fe11d8, fields=0x6fd5300, table=0x0, procedure=0x0) at
sql_select.cc:11140
#6 0x000000000060a479 in JOIN::exec
(this=0x6fe11d8) at sql_select.cc:2314
#7 0x00000000004d5102 in
subselect_single_select_engine::exec (this=0x6fdcda0) at
item_subselect.cc:1987
#8 0x00000000004d26b1 in Item_subselect::exec
(this=0x6fdccb0) at item_subselect.cc:280
#9 0x00000000004d1aaa in
Item_in_subselect::val_bool (this=0x2aaaabbc40b8) at
item_subselect.cc:880
#10 0x0000000000438821 in Item::val_bool_result
(this=0x2aaaabbc40b8) at item.h:745
#11 0x0000000000476941 in Item_in_optimizer::val_int
(this=0x6fe2a58) at item_cmpfunc.cc:1833
#12 0x00000000005e3d9a in evaluate_join_record (join=0x6fdce98,
join_tab=0x6fe3538, error=
out>) at sql_select.cc:11434
#13 0x00000000005e4215 in sub_select (join=0x6fdce98,
join_tab=0x6fe3538, end_of_records=
out>) at sql_select.cc:11391
#14 0x0000000000601d30 in do_select (join=0x6fdce98,
fields=0x6f819a0, table=0x0, procedure=0x0) at
sql_select.cc:11140
#15 0x000000000060a479 in JOIN::exec (this=0x6fdce98) at
sql_select.cc:2314
#16 0x000000000060ae0f in mysql_select (thd=0x6f7f980,
rref_pointer_array=0x6f81a68, tables=0x6fd4d90, wild_num=1,
fields=@0x6f819a0,
conds=0x6fdccb0, og_num=0, order=0x0, group=0x0, having=0x0,
proc_param=0x0, select_options=2147764736, result=0x6fdce78,
unit=0x6f81470,
select_lex=0x6f81898) at
sql_select.cc:2509
表 3build_template的bt
通过它我们可以看到该函数是由index_init调用的,而index_init又是在sub_select每次读取第一条记录的时候执行的,对于eq_ref则是调用join_read_key函数(注:e表,对于s表则调用join_read_always_key,它也有类似的过程)
static int
join_read_key(JOIN_TAB *tab)
{
int error;
TABLE *table= tab->table;
if
(!table->file->inited)
{
table->file->ha_index_init(tab->ref.key,
tab->sorted);
//在函数内部把table->file->inited赋值为INDEX;
}
…
可以看到table->file->inited这个变量决定着index_init的执行。那么它在哪里重新被赋值为0?答案就是ha_index_end(),而这个函数的执行是在:(再次bt。。。哈哈)
#0 st_join_table::cleanup (this=0x6fe28d0) at
handler.h:1186
#1 0x000000000060120b in JOIN::cleanup
(this=0x6fe0f10, full=
out>) at sql_select.cc:6982
#2 0x0000000000601734 in JOIN::join_free
(this=0x6fe0f10) at sql_select.cc:6905
#3 0x0000000000601db2
in do_select (join=0x6fe0f10,
fields=0x6f819a0, table=0x0, procedure=0x0) at
sql_select.cc:11161
#4 0x000000000060a479 in JOIN::exec
(this=0x6fe0f10) at sql_select.cc:2314
#5 0x000000000060ae0f in mysql_select
(thd=0x6f7f980, rref_pointer_array=0x6f81a68, tables=0x6fd5198,
wild_num=0, fields=@0x6f819a0,
conds=0x6fdd218, og_num=0, order=0x0, group=0x0, having=0x0,
proc_param=0x0, select_options=2147764736, result=0x6fdd398,
unit=0x6f81470,
select_lex=0x6f81898) at
sql_select.cc:2509
注:这里并没有显示这个函数名,但是handler.h:1186这条语句就是在ha_index_end()。我们看到这个函数是在do_select进行清除操作的时候调用的,而我们前面也已经说明了do_select是nest-loop的入口及出口地方,那么就是说每执行一交nest-loop的话都要进行一次index_init操作。对于in的类型,图4我们已经说明了它有count(t)的nest-loop过程,所以它就要执行443308次,这个刚才就等于gcov输出的join_read_key:table->file->ha_index_init(tab->ref.key,
tab->sorted);而join_read_always_key:table->file->ha_index_init(tab->ref.key,tab->sorted);总共执行了177224次,这个数字刚好是select
count(*) from titles t straight_join employees e
on(e.emp_no=t.emp_no) where e.gender='F';即每条满足t与e
join条件的记录。通过这两组数据,也验证了我们上面关于in的执行过程的描述。同样的这两个函数在join方式下都只执行了一次,因为它们只执行了一次nest-loop。
3. 总结
我们这里只是分析了占采样比例较高的build_template情况,对于其它的也是一样的分析方法,这里不再赘述。对于直接join方式与in子查询方式,两者select的时间复杂度是一样的(注:这里的select是指获得数据的方式,个数)。唯一不同的是对于in子查询它每次执行内部查询的时候都必须重新构造一个JOIN结构,完成相应的初始化操作,并且在这次内部查询结束之后,要完成相应的析构函数,如index_init,index_end,而当外部查询是全表扫描的时候这些操作的次数就是它的记录数,那么它们(构造,析构)所占用的性能也是显而易见的。简单一句话子查询的性能除了查询外,还消耗在JOIN的构造与析构过程,也就是上面表2的【t表】与【e表】中间的部分。