MySQL再深入执行计划之trace工具
Posted IT王小二
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了MySQL再深入执行计划之trace工具相关的知识,希望对你有一定的参考价值。
作者:IT王小二
不知不觉过年半个多月没更新啦,不得不说老家真冷啊,完全不想伸出小二滴爪子,回到深圳开始更新啦,还是深圳暖和!
上一篇 Explain执行计划详解 中提到执行计划分析查询语句时 possible_keys
列中存在可以使用的索引,但是实际最后 key
列中却并没有使用索引,走的全表扫描,这是为啥捏?
这就涉及小二说的 MySQL 执行查询语句时需要进行的查询成本分析了,而 MySQL 提供的 trace
工具就可以让我们看到 MySQL 计算的查询成本,以及选择索引的大致分析过程。
一、示例表
-- 示例表
CREATE TABLE `employees` (
`id` int(11) NOT NULL AUTO_INCREMENT,
`name` varchar(24) NOT NULL DEFAULT \'\' COMMENT \'姓名\',
`age` int(11) NOT NULL DEFAULT \'0\' COMMENT \'年龄\',
`position` varchar(20) NOT NULL DEFAULT \'\' COMMENT \'职位\',
`hire_time` timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP COMMENT \'入职时间\',
PRIMARY KEY (`id`),
KEY `idx_name_age_position` (`name`,`age`,`position`) USING BTREE
) ENGINE=InnoDB AUTO_INCREMENT=1 DEFAULT CHARSET=utf8 COMMENT=\'员工记录表\';
INSERT INTO employees(name,age,position,hire_time) VALUES(\'itwxe\',22,\'manager\',NOW());
INSERT INTO employees(name,age,position,hire_time) VALUES(\'weiwei\', 23,\'test\',NOW());
INSERT INTO employees(name,age,position,hire_time) VALUES(\'leilei\',23,\'dev\',NOW());
-- 插入10w条测试数据
drop procedure if exists insert_employees;
delimiter $$
create procedure insert_employees()
begin
declare i int;
set i = 1;
while(i <= 100000)do
insert into employees(name, age, position) values(CONCAT(\'itwxe\', i), rand() * 42 + 18, \'dev\');
set i = i + 1;
end while;
end$$
delimiter ;
call insert_employees();
二、trace使用步骤
首先需要说明的是开启 trace 工具会影响 MySQL 性能,所以只能临时分析 SQL 使用,用完之后立即关闭。
使用 trace 工具步骤:
-- 首先开启trace
mysql> set session optimizer_trace="enabled=on", end_markers_in_json=on;
-- 执行查询SQL
mysql> select * from employees where name > \'wei\' order by position;
-- 查询trace字段
mysql> SELECT * FROM information_schema.OPTIMIZER_TRACE;
使用完成后关闭 trace 工具:
-- 当分析完SQL,关闭trace
mysql> set session optimizer_trace="enabled=off";
三、trace字段解析
例如这条语句:
explain select * from employees where name > \'itwxe\' order by position;
explain select * from employees where name > \'wei\' order by position;
可以看到第一条 SQL 中 possible_keys
有可以使用的索引 idx_name_age_position
,但是最后实际上 MySQL 没有使用 idx_name_age_position
,而是使用了全表扫描;而第二条 SQL 中使用到了 idx_name_age_position
中的 name 列,这是为什么呢?
如果非常了解 MySQL 索引底层数据结构的小伙伴应该就知道因为第二条 SQL name > \'itwxe\'
扫描的联合索引 idx_name_age_position
的结果太多,回表成本太大,所以 MySQL 选择了全表扫描。要想看到 MySQL 计算的全表扫描成本和使用 idx_name_age_position
成本就得使用咱们这篇的主角 trace 工具了,以第二条SQL为例。
-- 首先开启trace
mysql> set session optimizer_trace="enabled=on", end_markers_in_json=on;
-- 执行查询SQL
mysql> select * from employees where name > \'wei\' order by position;
-- 查询trace字段
mysql> SELECT * FROM information_schema.OPTIMIZER_TRACE;
执行后可以得到如下的示例结果,重要字段注释如下:
| select * from employees where name > \'wei\' order by position |
"steps": [
"join_preparation": -- 第一阶段:SQL准备阶段,格式化SQL
"select#": 1,
"steps": [
"expanded_query": "/* select#1 */ select `employees`.`id` AS `id`,`employees`.`name` AS `name`,`employees`.`age` AS `age`,`employees`.`position` AS `position`,`employees`.`hire_time` AS `hire_time` from `employees` where (`employees`.`name` > \'wei\') order by `employees`.`position`"
] /* steps */
/* join_preparation */
,
"join_optimization": -- 第二阶段:SQL优化阶段
"select#": 1,
"steps": [
"condition_processing": -- 条件处理,where,1=1之类的SQL优化
"condition": "WHERE",
"original_condition": "(`employees`.`name` > \'wei\')",
"steps": [
"transformation": "equality_propagation",
"resulting_condition": "(`employees`.`name` > \'wei\')"
,
"transformation": "constant_propagation",
"resulting_condition": "(`employees`.`name` > \'wei\')"
,
"transformation": "trivial_condition_removal",
"resulting_condition": "(`employees`.`name` > \'wei\')"
] /* steps */
/* condition_processing */
,
"substitute_generated_columns":
/* substitute_generated_columns */
,
"table_dependencies": [ -- 表依赖详情
"table": "`employees`",
"row_may_be_null": false,
"map_bit": 0,
"depends_on_map_bits": [
] /* depends_on_map_bits */
] /* table_dependencies */
,
"ref_optimizer_key_uses": [
] /* ref_optimizer_key_uses */
,
"rows_estimation": [ -- 预估表的访问成本
"table": "`employees`",
"range_analysis":
"table_scan": -- 全表扫描情况
"rows": 100075, -- 扫描行数,innodb中这个是预估值,MyISAM中时准确值
"cost": 20306 -- 查询成本
/* table_scan */,
"potential_range_indexes": [ -- 查询可能使用的索引
"index": "PRIMARY", -- 主键索引
"usable": false,
"cause": "not_applicable"
,
"index": "idx_name_age_position", -- 联合索引
"usable": true,
"key_parts": [
"name",
"age",
"position",
"id"
] /* key_parts */
] /* potential_range_indexes */,
"setup_range_conditions": [
] /* setup_range_conditions */,
"group_index_range":
"chosen": false,
"cause": "not_group_by_or_distinct"
/* group_index_range */,
"analyzing_range_alternatives": -- 分析各个索引使用成本
"range_scan_alternatives": [
"index": "idx_name_age_position",
"ranges": [
"wei < name" -- 索引使用范围
] /* ranges */,
"index_dives_for_eq_ranges": true,
"rowid_ordered": false, -- 使用该索引获取的记录是否按照主键排序
"using_mrr": false, -- 是否使用mrr
"index_only": false, -- 是否使用覆盖索引
"rows": 1, -- 索引扫描行数,也是个预估值
"cost": 2.21, -- 索引使用成本
"chosen": true -- 是否选择该索引
] /* range_scan_alternatives */,
"analyzing_roworder_intersect":
"usable": false,
"cause": "too_few_roworder_scans"
/* analyzing_roworder_intersect */
/* analyzing_range_alternatives */,
"chosen_range_access_summary":
"range_access_plan":
"type": "range_scan",
"index": "idx_name_age_position",
"rows": 1,
"ranges": [
"wei < name"
] /* ranges */
/* range_access_plan */,
"rows_for_plan": 1,
"cost_for_plan": 2.21,
"chosen": true
/* chosen_range_access_summary */
/* range_analysis */
] /* rows_estimation */
,
"considered_execution_plans": [
"plan_prefix": [
] /* plan_prefix */,
"table": "`employees`",
"best_access_path": -- 最优访问路径
"considered_access_paths": [ -- 最终访问路径
"rows_to_scan": 1,
"access_type": "range", -- 访问类型 range:访问扫描;scan:全表扫描
"range_details":
"used_index": "idx_name_age_position"
/* range_details */,
"resulting_rows": 1,
"cost": 2.41,
"chosen": true, -- 确定选择
"use_tmp_table": true
] /* considered_access_paths */
/* best_access_path */,
"condition_filtering_pct": 100,
"rows_for_plan": 1,
"cost_for_plan": 2.41,
"sort_cost": 1,
"new_cost_for_plan": 3.41,
"chosen": true
] /* considered_execution_plans */
,
"attaching_conditions_to_tables":
"original_condition": "(`employees`.`name` > \'wei\')",
"attached_conditions_computation": [
] /* attached_conditions_computation */,
"attached_conditions_summary": [
"table": "`employees`",
"attached": "(`employees`.`name` > \'wei\')"
] /* attached_conditions_summary */
/* attaching_conditions_to_tables */
,
"clause_processing":
"clause": "ORDER BY",
"original_clause": "`employees`.`position`",
"items": [
"item": "`employees`.`position`"
] /* items */,
"resulting_clause_is_simple": true,
"resulting_clause": "`employees`.`position`"
/* clause_processing */
,
"reconsidering_access_paths_for_index_ordering":
"clause": "ORDER BY",
"steps": [
] /* steps */,
"index_order_summary":
"table": "`employees`",
"index_provides_order": false,
"order_direction": "undefined",
"index": "idx_name_age_position",
"plan_changed": false
/* index_order_summary */
/* reconsidering_access_paths_for_index_ordering */
,
"refine_plan": [
"table": "`employees`",
"pushed_index_condition": "(`employees`.`name` > \'wei\')",
"table_condition_attached": null
] /* refine_plan */
] /* steps */
/* join_optimization */
,
"join_execution": -- 第三阶段:SQL执行阶段
"select#": 1,
"steps": [
"filesort_information": [
"direction": "asc",
"table": "`employees`",
"field": "position"
] /* filesort_information */,
"filesort_priority_queue_optimization":
"usable": false,
"cause": "not applicable (no LIMIT)"
/* filesort_priority_queue_optimization */,
"filesort_execution": [
] /* filesort_execution */,
"filesort_summary":
"rows": 1,
"examined_rows": 1,
"number_of_tmp_files": 0,
"sort_buffer_size": 262056,
"sort_mode": "<sort_key, packed_additional_fields>"
/* filesort_summary */
] /* steps */
/* join_execution */
] /* steps */
| 0 | 0 |
可以看到预估表的访问成本计算过程中,MySQL 会计算全表扫描和可以使用的索引成本,比较各个可以使用的索引和全表扫描的查询成本,最终选择最优的一种方式去查询。例如这个例子中全表扫描成本为20306,使用 idx_name_age_position
联合索引成本为2.21,所以最终 MySQL 选择使用 idx_name_age_position
索引去执行查询。
当然 trace 工具主要看的是 MySQL 大致的优化和计算成本过程,那么 MySQL 是通过什么去计算出的这个成本呢?小二先在这里挖个坑吧,看下啥时候填上,哈哈哈~~~
以上是关于MySQL再深入执行计划之trace工具的主要内容,如果未能解决你的问题,请参考以下文章
mysql 执行计划分析三看, explain,profiling,optimizer_trace