深入解析:从源码窥探MySQL优化器

深入解析:从源码窥探MySQL优化器

南川网站建设公司创新互联,南川网站设计制作,有大型网站制作公司丰富经验。已为南川上1000家提供企业网站建设服务。企业网站搭建\成都外贸网站制作要多少钱,请找那个售后服务好的南川做网站的公司定做!

作者 | 汤爱中,云和恩墨SQM开发者,Oracle/MySQL/DB2的SQL解析引擎、SQL审核与智能优化引擎的重要贡献者,产品广泛应用于金融、电信等行业客户中。

摘要

优化器是逻辑SQL到物理存储的解释器,是一个复杂而“愚蠢”的数学模型,它的入参通常是SQL、统计信息以及优化器参数等,而输出通常一个可执行的查询计划,因此优化器的优劣取决于数学模型的稳定性和健壮性,理解这个数学模型就能理解数据库的SQL处理流程。

01 优化器的执行流程 

深入解析:从源码窥探MySQL优化器

(注:此图出自李海翔)

上图展示了优化器的大致执行过程,可以简单描述为:

1 根据语法树及统计统计,构建初始表访问数组(init_plan_arrays)

2 根据表访问数组,计算每个表的最佳访问路径(find_best_ref),同时保存当前最优执行计划(COST最小)

3 如果找到更优的执行计划则更新最优执行计划,否则优化结束。

从上述流程可以看出,执行计划的生成是一个“动态规划/贪心算法”的过程,动态规划公式可以表示为:Min(Cost(Tn+1)) = Min(Cost(T1))+Min(Cost(T2))+...Min(Cost(Tn-1))+Min(Cost(Tn)),其中Cost(Tn)表示访问表T1 T2一直到Tn的代价。如果优化器没有任何先验知识,则需要进行 A(n,n) 次循环,是一个全排列过程,很显然优化器是有先验知识的,如表大小,外连接,子查询等都会使得表的访问是部分有序的,可以理解为一个“被裁减”的动态规划,动态规则的核心函数为:Join::Best_extention_limited_search,在源码中有如下代码结构:

bool Optimize_table_order::best_extension_by_limited_search(
         table_map remaining_tables,
         uint      idx,
         uint      current_search_depth)
{
      for (JOIN_TAB **pos= join->best_ref + idx; *pos; pos++)
    {
      ......
         best_access_path(s, remaining_tables, idx, false,
                       idx ? (position-1)->prefix_rowcount : 1.0,
                       position);
......
  if (best_extension_by_limited_search(remaining_tables_after,
                                             idx + 1,
                                             current_search_depth - 1))
          ......
          backout_nj_state(remaining_tables, s);
......
}
}

以上代码是在一个for循环中递归搜索,这是一个典型的全排列的算法。

02优化器参数

MySQL的优化器对于Oracle来说还显得比较幼稚。Oracle有着各种丰富的统计信息,比如系统统计信息,表统计信息,索引统计信息等,而MySQL则需要更多的常量,其中MySQL5.7提供了表mysql.server_cost和表mysql.engine_cost,可以供用户配置,使得用户能够调整优化器模型,下面就几个常见而又非常重要的参数进行介绍:

1 #define ROW_EVALUATE_COST 0.2f

  计算符合条件的行的代价,行数越多,代价越大

2 #define IO_BLOCK_READ_COST 1.0f

  从磁盘读取一个Page的代价

3 #define MEMORY_BLOCK_READ_COST 1.0f

 从内存读取一个Page的代价,对于Innodb来说,表示从一个Buffer Pool读取一个Page的代价,因此读取内存页和磁盘页的默认代价是一样的

4 #define COND_FILTER_EQUALITY 0.1f

等值过滤条件默认值为0.1,例如name = ‘lily’, 表大小为100,则返回10行数据

5 #define COND_FILTER_INEQUALITY 0.3333f

非等值过滤条件的默认值是0.3333,例如col1>col2

6 #define COND_FILTER_BETWEEN 0.1111f

 Between过滤的默认值是0.1111f,例如:col1 between a and b

......

这样的常量很多,涉及到过滤条件、JOIN缓存、临时表等等各种代价,理解这些常量后,看到执行计划的Cost后,你会有种豁然开朗的感觉!

03 优化器选项

在MySQL中,执行select @@optimizer_trace, 得到如下参数:

index_merge=on,index_merge_union=off,index_merge_sort_union=off, index_merge_intersection=on, engine_condition_pushdown=on, index_condition_pushdown=on, mrr=on,mrr_cost_based=on,block_nested_loop=on,batched_key_access=off,materialization=on,semijoin=on,loosescan=on,firstmatch=on,subquery_materialization_cost_based=on, use_index_extensions=on, condition_fanout_filter=on

04 Optimize Trace是如何生成的?

在流程图中的函数中,存在大量如下代码:

Opt_trace_object trace_ls(trace, "searching_loose_scan_index");

因此,在优化器运行过程中,优化器的执行路径也被保存在Opt_trace_object中,进而保存在information_schema.optimizer_trace中,方便用户查询和跟踪。

05 优化器的典型使用场景

5.1 全表扫描

select * from sakila.actor;

表actor统计信息如下:

Db_name

Table_name

Last_update

n_rows

Cluster_index_size

Other_index

sakila

actor

2018-11-20 16:20:12

200

1

0

 

主键actor_id统计信息如下:

Index_name

Last_update

Stat_name

Stat_value

Sample_size

Stat_description

PRIMARY

2018-11-14 14:25:49

n_diff_pfx01

200

1

actor_id

PRIMARY

2018-11-14 14:25:49

n_leaf_pages

1

NULL

Number of leaf pages in the index

PRIMARY

2018-11-14 14:25:49

size

1

NULL

Number of pages in the index

执行计划:  

{                                                                                           
  "query_block": {                                                                                        
    "select_id": 1,                                                                       
    "cost_info": {                                                                                            
      "query_cost": "41.00"                                                                                    
    },                                                       
    "table": {                                                            
      "table_name": "actor",                                                                                    
      "access_type": "ALL",                                                                      
      "rows_examined_per_scan": 200,                                                                 
      "rows_produced_per_join": 200,                              
      "filtered": "100.00",                                                                  
      "cost_info": {                                                          
        "read_cost": "1.00",                                                                    
        "eval_cost": "40.00",                                                              
        "prefix_cost": "41.00",                                                               
        "data_read_per_join": "56K"  
      },                                                       
      "used_columns": [                                                                 
        "actor_id",                                                      
        "first_name",                                                            
        "last_name",                                                             
        "last_update",                                                            
        "id"                                                          
      ]                                                         
    }                                                                 
  }                                                                
}
IO_COST = CLUSTER_INDEX_SIZE * PAGE_READ_TIME = 1 * 1 =1;
EVAL_COST = TABLE_ROWS*EVALUATE_COST = 200 * 0.2 =40;
PREFIX_COST = IO_COST + EVAL_COST;

注意以上过程忽略了内存页和磁盘页的访问代价差异。

 5.2 表连接时使用全表扫描

SELECT
  *
FROM
  sakila.actor a,
  sakila.film_actor b
WHERE a.actor_id = b.actor_id

Db_name

Table_name

Last_update

n_rows

Cluster_index_size

Other_index_size

Sakila

Film_actor

2018-11-20 16:55:31

5462

12

5

表film_actor中索引(actor_id,film_id)统计信息如下:

Index_name

Last_update

Stat_name

Stat_value

Sample_size

Stat_description

PRIMARY

2018-11-14 14:25:49

n_diff_pfx01

200

1

actor_id

PRIMARY

2018-11-14 14:25:49

n_diff_pfx02

5462

1

actor_id,film_id

PRIMARY

2018-11-14 14:25:49

n_leaf_pages

11

NULL

Number of leaf pages in the index

PRIMARY

2018-11-14 14:25:49

size

12

NULL

Number of pages in theindex

{                                      
  "query_block": {                   
    "select_id": 1,                    
    "cost_info": {                   
      "query_cost": "1338.07"             
    },                                
    "nested_loop": [                
      {                              
        "table": {       
          "table_name": "a",  
          "access_type": "ALL", 
          "possible_keys": [   
            "PRIMARY"            
          ],                  
          "rows_examined_per_scan": 200, 
          "rows_produced_per_join": 200, 
          "filtered": "100.00", 
          "cost_info": {       
            "read_cost": "1.00", 
            "eval_cost": "40.00", 
            "prefix_cost": "41.00",  
            "data_read_per_join": "54K"  
          },                     
          "used_columns": [     
            "actor_id",      
            "first_name",    
            "last_name",      
            "last_update"          
          ]                          
        }                              
      },                              
      {                            
        "table": {               
          "table_name": "b",     
          "access_type": "ref",  
          "possible_keys": [  
            "PRIMARY"               
          ],                     
          "key": "PRIMARY",       
          "used_key_parts": [     
            "actor_id"              
          ],                    
          "key_length": "2",     
          "ref": [            
            "sakila.a.actor_id"    
          ],                 
          "rows_examined_per_scan": 27,  
          "rows_produced_per_join": 5461, 
          "filtered": "100.00",     
          "cost_info": {          
            "read_cost": "204.67", 
            "eval_cost": "1092.40", 
            "prefix_cost": "1338.07", 
            "data_read_per_join": "85K" 
          }, 
          "used_columns": [  
            "actor_id",            
            "film_id",        
            "last_update"           
          ]                              
        }                               
      }                                 
    ]                                     
  }                                          
}

第一张表actor的全表扫代价为41,可以参考示例1。

        第二个表就是NET LOOP 代价:

read_cost(204.67) =prefix_rowcount * (1 + keys_per_value/table_rows*cluster_index_size =

200 * (1+27/13863*12)*1

注意:27 相当于对于每个actor_id,film_actor的索引估计,对于每个actor_id,平均有27条记录=5462/200

Table_rows是如何计算的呢?

Film_actor表的实际记录数是5462,一共12个page,11个叶子页,总大小为11*16K(默认页大小)=180224Byte, 最小记录长度为26(通过计算字段长度可得),13863 = 180224/26*2, 2是安全因子,做最差的代价估计。

表连接返回行数=200*5462/200,因此行估算代价为5462*0.2=1902.4

5.3 IN查询

表film_actor中索引idx_id(film_id)统计信息如下:

Index_name

Last_update

Stat_name

Stat_value

Sample_size

Stat_description

idx_id

2018-11-14 14:25:49

n_diff_pfx01

997

4

actor_id

idx_id

2018-11-14 14:25:49

n_diff_pfx02

5462

4

film_id,actor_id

idx_id

2018-11-14 14:25:49

n_leaf_pages

4

NULL

Number of leaf pages in the index

idx_id

2018-11-14 14:25:49

size

5

NULL

Number of pages in the index

EXPLAIN SELECT * FROM ACTOR WHERE actor_id IN (SELECT film_id FROM film_actor)
{                                  
  "query_block": {              
    "select_id": 1,             
    "cost_info": {                  
      "query_cost": "460.79"    
    },                            
    "nested_loop": [               
      {                      
        "table": {        
          "table_name": "ACTOR", 
          "access_type": "ALL", 
          "possible_keys": [ 
            "PRIMARY"          
          ],               
          "rows_examined_per_scan": 200, 
          "rows_produced_per_join": 200, 
          "filtered": "100.00", 
          "cost_info": { 
            "read_cost": "1.00", 
            "eval_cost": "40.00", 
            "prefix_cost": "41.00", 
            "data_read_per_join": "56K" 
          },                 
          "used_columns": [ 
            "actor_id",  
            "first_name",   
            "last_name", 
            "last_update",   
            "id"                 
          ]                    
        }                         
      },                         
      {                        
        "table": {     
          "table_name": "film_actor",     
          "access_type": "ref", 
          "possible_keys": [ 
            "idx_id"            
          ],                
          "key": "idx_id", 
          "used_key_parts": [ 
            "film_id"           
          ],               
          "key_length": "2",   
          "ref": [      
            "sakila.ACTOR.actor_id" 
          ],           
          "rows_examined_per_scan": 5, 
          "rows_produced_per_join": 200, 
          "filtered": "100.00", 
          "using_index": true, 
          "first_match": "ACTOR", 
          "cost_info": { 
            "read_cost": "200.66", 
            "eval_cost": "40.00", 
            "prefix_cost": "460.79", 
            "data_read_per_join": "3K" 
          },                      
          "used_columns": [  
            "film_id"            
          ],                      
          "attached_condition": "(`sakila`.`actor`.`actor_id` = `sakila`.`film_actor`.`film_id`)"                                                                        
        }                            
      }                                
    ]                                     
  }                                       
}
id  select_type  table       partitions  type    possible_keys  key     key_len  ref                      rows  filtered  Extra                                        
------  -----------  ----------  ----------  ------  -------------  ------  -------  ---------------------  ------  --------  ---------------------------------------------
     1  SIMPLE       ACTOR       (NULL)      ALL     PRIMARY        (NULL)  (NULL)   (NULL)                    200    100.00  (NULL)                                       
     1  SIMPLE       film_actor  (NULL)      ref     idx_id         idx_id  2 
sakila.ACTOR.actor_id       5    100.00  Using where; Using index; FirstMatch(ACTOR)

从执行计划中可以看出,MySQL采用FirstMatch方式。在MySQL中,半链接优化方式为:Materialization Strategy,LooseScan,FirstMatch,DuplicateWeedout,默认情况下四种优化方式都是存在的,选取方式基于最小COST。现在我们以FirstMatch为例,讲解优化器的执行流程。

 SQL如下:

  select * from Country
where Country.code IN (select City.Country
                       from City
                       where City.Population > 1*1000*1000)
      and Country.continent='Europe'

深入解析:从源码窥探MySQL优化器

从上图可以看出,FirstMatch是通过判断记录是否已经在结果集中存在来减少查询和匹配流程。

表actor的访问代价可以参考示例1.

表film_actor表的访问代价200.66是如何计算的呢?

访问表film_actor中索引字段film_id,MySQL会走覆盖索引扫,即IDEX_ONLY_SCAN,一次索引访问的代价是如何计算的呢?

参考函数double handler::index_only_read_time(uint keynr, double records)

索引块大小为16K,并且MySQL假设块都是半满的,则一个块能够存放的索引记录数为:

16K/2/(索引长度+主键长度(注:二级索引存储的是主键的引用))=16K/2/(2+4)+1=1366,

其中主键为(actor_id,film_id),两个字段都是smallint,占用4个字节,而索引idx_id(film_id)是2个字节,因此每次访问索引的代价为:(5.47+1366-1)/1366 = 1.0032, 访问film_actor表一共需要200次,总访问代价为:200*1.0032=200.66

总代价460.79 = 表actor的访问代价+表film_actor访问代价+行估算代价=

41+200.66+200*1*5.47*1*02,其中两个1分别表示过滤因子,由于两个表均没有过滤条件因此过滤因子都是1。


分享标题:深入解析:从源码窥探MySQL优化器
文章起源:http://scyanting.com/article/jgieoe.html