MySQL 调优 | OPTIMIZER_TRACE详解
TIPS
本文基于MySQL 8.0编写,理论支持MySQL 5.6及更高版本。
OPTIMIZER_TRACE是MySQL 5.6引入的一项跟踪功能,它可以跟踪优化器做出的各种决策(比如访问表的方法、各种开销计算、各种转换等),并将跟踪结果记录到 INFORMATION_SCHEMA.OPTIMIZER_TRACE
表中。此功能默认关闭,开启后,可分析如下语句:
- SELECT
- INSERT
- REPLACE
- UPDATE
- DELETE
- EXPLAIN
- SET
- DECLARE
- CASE
- IF
- RETURN
- CALL
OPTIMIZER_TRACE相关参数
TIPS
参考 https://dev.mysql.com/doc/internals/en/system-variables-controlling-trace.html
-
optimizer_trace
- optimizer_trace总开关,默认值:
enabled=off,one_line=off
- enabled:是否开启optimizer_trace;on表示开启,off表示关闭。
- one_line:是否开启单行存储。on表示开启;off表示关闭,将会用标准的JSON格式化存储。设置成on将会有良好的格式,设置成off可节省一些空间。
- optimizer_trace总开关,默认值:
-
optimizer_trace_features
- 控制optimizer_trace跟踪的内容,默认值:
greedy_search=on,range_optimizer=on,dynamic_range=on,repeated_subselect=on
,表示开启所有跟踪项。
- 控制optimizer_trace跟踪的内容,默认值:
-
greedy_search:是否跟踪贪心搜索,有关贪心算法详见 https://blog.csdn.net/weixin_42813521/article/details/105563103
- range_optimizer:是否跟踪范围优化器
-
dynamic_range:是否跟踪动态范围优化
- repeated_subselect:是否跟踪子查询,如果设置成off,只跟踪第一条Item_subselect的执行
-
详见 https://dev.mysql.com/doc/internals/en/optimizer-features-to-trace.html
-
optimizer_trace_limit:控制optimizer_trace展示多少条结果,默认1
-
optimizer_trace_max_mem_size:optimizer_trace堆栈信息允许的最大内存,默认1048576
-
optimizer_trace_offset:第一个要展示的optimizer trace的偏移量,默认-1。
-
end_markers_in_json:如果JSON结构很大,则很难将右括号和左括号配对。为了帮助读者阅读,可将其设置成on,这样会在右括号附近加上注释,默认off。
参考: https://dev.mysql.com/doc/internals/en/end-markers-in-json-system-variable.html
TIPS
以上参数可用SET语句操作,例如,用如下命令即可打开OPTIMIZER TRACE
> SET OPTIMIZER_TRACE="enabled=on",END_MARKERS_IN_JSON=on; >
也可用SET GLOBAL全局开启。但即使全局开启OPTIMIZER_TRACE,每个Session也只能跟踪它自己执行的语句:
> SET GLOBAL OPTIMIZER_TRACE="enabled=on",END_MARKERS_IN_JSON=on; >
optimizer_trace_limit和optimizer_trace_offset这两个参数经常配合使用,例如:
> SET optimizer_trace_offset=<OFFSET>, optimizer_trace_limit=<LIMIT> >
这两个参数配合使用,有点类似MySQL里面的 limit语句。
默认情况下,由于optimizer_trace_offset=-1,optimizer_trace_limit=1,记录最近的一条SQL语句,展示时,每次展示1条数据;
如果改成
SET optimizer_trace_offset=-2, optimizer_trace_limit=1
,则会记录倒数第二条SQL语句;有关 optimizer_trace_offset 、optimizer_trace_limit更多细节,可参考 https://dev.mysql.com/doc/internals/en/tuning-trace-purging.html
OPTIMIZER_TRACE使用
-
开启OPTIMIZER_TRACE功能,并设置要展示的数据条目数:
SET OPTIMIZER_TRACE="enabled=on",END_MARKERS_IN_JSON=on; SET optimizer_trace_offset=-30, optimizer_trace_limit=30;
-
发送你想要分析的SQL语句,例如:
select * from salaries where from_date = '1986-06-26' and to_date = '1987-06-26';
-
使用如下语句分析,即可获得类似如下的结果:
mysql> SELECT * FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE limit 30 \G; *************************** 1. row *************************** QUERY: select * from salaries where from_date = '1986-06-26' and to_date = '1987-06-26' TRACE: { "steps": [ { "join_preparation": { "select#": 1, "steps": [ { "expanded_query": "/* select#1 */ select `salaries`.`emp_no` AS `emp_no`,`salaries`.`salary` AS `salary`,`salaries`.`from_date` AS `from_date`,`salaries`.`to_date` AS `to_date` from `salaries` where ((`salaries`.`from_date` = '1986-06-26') and (`salaries`.`to_date` = '1987-06-26'))" } ] /* steps */ } /* join_preparation */ }, { "join_optimization": { "select#": 1, "steps": [ { "condition_processing": { "condition": "WHERE", "original_condition": "((`salaries`.`from_date` = '1986-06-26') and (`salaries`.`to_date` = '1987-06-26'))", "steps": [ { "transformation": "equality_propagation", "resulting_condition": "(multiple equal('1986-06-26', `salaries`.`from_date`) and multiple equal('1987-06-26', `salaries`.`to_date`))" }, { "transformation": "constant_propagation", "resulting_condition": "(multiple equal('1986-06-26', `salaries`.`from_date`) and multiple equal('1987-06-26', `salaries`.`to_date`))" }, { "transformation": "trivial_condition_removal", "resulting_condition": "(multiple equal(DATE'1986-06-26', `salaries`.`from_date`) and multiple equal(DATE'1987-06-26', `salaries`.`to_date`))" } ] /* steps */ } /* condition_processing */ }, { "substitute_generated_columns": { } /* substitute_generated_columns */ }, { "table_dependencies": [ { "table": "`salaries`", "row_may_be_null": false, "map_bit": 0, "depends_on_map_bits": [ ] /* depends_on_map_bits */ } ] /* table_dependencies */ }, { "ref_optimizer_key_uses": [ { "table": "`salaries`", "field": "from_date", "equals": "DATE'1986-06-26'", "null_rejecting": false }, { "table": "`salaries`", "field": "to_date", "equals": "DATE'1987-06-26'", "null_rejecting": false } ] /* ref_optimizer_key_uses */ }, { "rows_estimation": [ { "table": "`salaries`", "range_analysis": { "table_scan": { "rows": 2838216, "cost": 286799 } /* table_scan */, "potential_range_indexes": [ { "index": "PRIMARY", "usable": false, "cause": "not_applicable" }, { "index": "salaries_from_date_to_date_index", "usable": true, "key_parts": [ "from_date", "to_date", "emp_no" ] /* 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 */, "skip_scan_range": { "potential_skip_scan_indexes": [ { "index": "salaries_from_date_to_date_index", "usable": false, "cause": "query_references_nonkey_column" } ] /* potential_skip_scan_indexes */ } /* skip_scan_range */, "analyzing_range_alternatives": { "range_scan_alternatives": [ { "index": "salaries_from_date_to_date_index", "ranges": [ "0xda840f <= from_date <= 0xda840f AND 0xda860f <= to_date <= 0xda860f" ] /* ranges */, "index_dives_for_eq_ranges": true, "rowid_ordered": true, "using_mrr": false, "index_only": false, "rows": 86, "cost": 50.909, "chosen": true } ] /* 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": "salaries_from_date_to_date_index", "rows": 86, "ranges": [ "0xda840f <= from_date <= 0xda840f AND 0xda860f <= to_date <= 0xda860f" ] /* ranges */ } /* range_access_plan */, "rows_for_plan": 86, "cost_for_plan": 50.909, "chosen": true } /* chosen_range_access_summary */ } /* range_analysis */ } ] /* rows_estimation */ }, { "considered_execution_plans": [ { "plan_prefix": [ ] /* plan_prefix */, "table": "`salaries`", "best_access_path": { "considered_access_paths": [ { "access_type": "ref", "index": "salaries_from_date_to_date_index", "rows": 86, "cost": 50.412, "chosen": true }, { "access_type": "range", "range_details": { "used_index": "salaries_from_date_to_date_index" } /* range_details */, "chosen": false, "cause": "heuristic_index_cheaper" } ] /* considered_access_paths */ } /* best_access_path */, "condition_filtering_pct": 100, "rows_for_plan": 86, "cost_for_plan": 50.412, "chosen": true } ] /* considered_execution_plans */ }, { "attaching_conditions_to_tables": { "original_condition": "((`salaries`.`to_date` = DATE'1987-06-26') and (`salaries`.`from_date` = DATE'1986-06-26'))", "attached_conditions_computation": [ ] /* attached_conditions_computation */, "attached_conditions_summary": [ { "table": "`salaries`", "attached": "((`salaries`.`to_date` = DATE'1987-06-26') and (`salaries`.`from_date` = DATE'1986-06-26'))" } ] /* attached_conditions_summary */ } /* attaching_conditions_to_tables */ }, { "finalizing_table_conditions": [ { "table": "`salaries`", "original_table_condition": "((`salaries`.`to_date` = DATE'1987-06-26') and (`salaries`.`from_date` = DATE'1986-06-26'))", "final_table_condition ": null } ] /* finalizing_table_conditions */ }, { "refine_plan": [ { "table": "`salaries`" } ] /* refine_plan */ } ] /* steps */ } /* join_optimization */ }, { "join_execution": { "select#": 1, "steps": [ ] /* steps */ } /* join_execution */ } ] /* steps */ } MISSING_BYTES_BEYOND_MAX_MEM_SIZE: 0 INSUFFICIENT_PRIVILEGES: 0 1 row in set (0.00 sec)
-
分析完成,关闭OPTIMIZER_TRACE
SET optimizer_trace="enabled=off";
OPTIMIZER_TRACE结果分析
由上面的结果可知,OPTIMIZER_TRACE有四个字段:
- QUERY:查询语句
- TRACE:QUERY字段对应语句的跟踪信息
- MISSING_BYTES_BEYOND_MAX_MEM_SIZE:跟踪信息过长时,被截断的跟踪信息的字节数。
- INSUFFICIENT_PRIVILEGES:执行跟踪语句的用户是否有查看对象的权限。当不具有权限时,该列信息为1且TRACE字段为空,一般在调用带有SQL SECURITY DEFINER的视图或者是存储过程的情况下,会出现此问题。
TIPS
参考: https://dev.mysql.com/doc/refman/8.0/en/optimizer-trace-table.html
最核心的是TRACE字段的内容。我们逐段分析:
join_preparation
join_preparation段落展示了准备阶段的执行过程。
{
"join_preparation": {
"select#": 1,
"steps": [
{
-- 对比下原始语句,可以知道,这一步做了个格式化。
"expanded_query": "/* select#1 */ select `salaries`.`emp_no` AS `emp_no`,`salaries`.`salary` AS `salary`,`salaries`.`from_date` AS `from_date`,`salaries`.`to_date` AS `to_date` from `salaries` where ((`salaries`.`from_date` = '1986-06-26') and (`salaries`.`to_date` = '1987-06-26'))"
}
]
/* steps */
}
/* join_preparation */
}
join_optimization
join_optimization展示了优化阶段的执行过程,是分析OPTIMIZER TRACE的重点。这段内容超级长,而且分了好多步骤,不妨按照步骤逐段分析:
condition_processing
该段用来做条件处理,主要对WHERE条件进行优化处理。
"condition_processing": {
"condition": "WHERE",
"original_condition": "((`salaries`.`from_date` = '1986-06-26') and (`salaries`.`to_date` = '1987-06-26'))",
"steps": [
{
"transformation": "equality_propagation",
"resulting_condition": "(multiple equal('1986-06-26', `salaries`.`from_date`) and multiple equal('1987-06-26', `salaries`.`to_date`))"
},
{
"transformation": "constant_propagation",
"resulting_condition": "(multiple equal('1986-06-26', `salaries`.`from_date`) and multiple equal('1987-06-26', `salaries`.`to_date`))"
},
{
"transformation": "trivial_condition_removal",
"resulting_condition": "(multiple equal(DATE'1986-06-26', `salaries`.`from_date`) and multiple equal(DATE'1987-06-26', `salaries`.`to_date`))"
}
] /* steps */
} /* condition_processing */
其中:
- condition:优化对象类型。WHERE条件句或者是HAVING条件句
- original_condition:优化前的原始语句
- steps:主要包括三步,分别是quality_propagation(等值条件句转换),constant_propagation(常量条件句转换),trivial_condition_removal(无效条件移除的转换)
- transformation:转换类型句
- resulting_condition:转换之后的结果输出
substitute_generated_columns
substitute_generated_columns用于替换虚拟生成列
"substitute_generated_columns": {
} /* substitute_generated_columns */
table_dependencies
分析表之间的依赖关系
{
"table_dependencies": [
{
"table": "`salaries`",
"row_may_be_null": false,
"map_bit": 0,
"depends_on_map_bits": [
] /* depends_on_map_bits */
}
] /* table_dependencies */
}
其中:
- table:涉及的表名,如果有别名,也会展示出来
- row_may_be_null:行是否可能为NULL,这里是指JOIN操作之后,这张表里的数据是不是可能为NULL。如果语句中使用了LEFT JOIN,则后一张表的row_may_be_null会显示为true
- map_bit:表的映射编号,从0开始递增
- depends_on_map_bits:依赖的映射表。主要是当使用STRAIGHT_JOIN强行控制连接顺序或者LEFT JOIN/RIGHT JOIN有顺序差别时,会在depends_on_map_bits中展示前置表的map_bit值。
ref_optimizer_key_uses
列出所有可用的ref类型的索引。如果使用了组合索引的多个部分(例如本例,用到了index(from_date, to_date) 的多列索引),则会在ref_optimizer_key_uses下列出多个元素,每个元素中会列出ref使用的索引及对应值。
{
"ref_optimizer_key_uses": [
{
"table": "`salaries`",
"field": "from_date",
"equals": "DATE'1986-06-26'",
"null_rejecting": false
},
{
"table": "`salaries`",
"field": "to_date",
"equals": "DATE'1987-06-26'",
"null_rejecting": false
}
] /* ref_optimizer_key_uses */
}
rows_estimation
顾名思义,用于估算需要扫描的记录数。
{
"rows_estimation": [
{
"table": "`salaries`",
"range_analysis": {
"table_scan": {
"rows": 2838216,
"cost": 286799
} /* table_scan */,
"potential_range_indexes": [
{
"index": "PRIMARY",
"usable": false,
"cause": "not_applicable"
},
{
"index": "salaries_from_date_to_date_index",
"usable": true,
"key_parts": [
"from_date",
"to_date",
"emp_no"
] /* 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 */,
"skip_scan_range": {
"potential_skip_scan_indexes": [
{
"index": "salaries_from_date_to_date_index",
"usable": false,
"cause": "query_references_nonkey_column"
}
] /* potential_skip_scan_indexes */
} /* skip_scan_range */,
"analyzing_range_alternatives": {
"range_scan_alternatives": [
{
"index": "salaries_from_date_to_date_index",
"ranges": [
"0xda840f <= from_date <= 0xda840f AND 0xda860f <= to_date <= 0xda860f"
] /* ranges */,
"index_dives_for_eq_ranges": true,
"rowid_ordered": true,
"using_mrr": false,
"index_only": false,
"rows": 86,
"cost": 50.909,
"chosen": true
}
] /* 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": "salaries_from_date_to_date_index",
"rows": 86,
"ranges": [
"0xda840f <= from_date <= 0xda840f AND 0xda860f <= to_date <= 0xda860f"
] /* ranges */
} /* range_access_plan */,
"rows_for_plan": 86,
"cost_for_plan": 50.909,
"chosen": true
} /* chosen_range_access_summary */
} /* range_analysis */
}
] /* rows_estimation */
}
其中:
-
table:表名
-
range_analysis:
-
table_scan:如果全表扫描的话,需要扫描多少行(row,2838216),以及需要的代价(cost,286799)
-
potential_range_indexes:列出表中所有的索引并分析其是否可用。如果不可用的话,会列出不可用的原因是什么;如果可用会列出索引中可用的字段;
-
setup_range_conditions:如果有可下推的条件,则带条件考虑范围查询
-
group_index_range:当使用了GROUP BY或DISTINCT时,是否有合适的索引可用。当未使用GROUP BY或DISTINCT时,会显示chosen=false, cause=not_group_by_or_distinct;如使用了GROUP BY或DISTINCT,但是多表查询时,会显示chosen=false,cause =not_single_table。其他情况下会尝试分析可用的索引(potential_group_range_indexes)并计算对应的扫描行数及其所需代价
-
skip_scan_range:是否使用了skip scan
TIPS
skip_scan_range是MySQL 8.0的新特性,感兴趣的可详见 https://blog.csdn.net/weixin_43970890/article/details/89494915
-
analyzing_range_alternatives:分析各个索引的使用成本
- range_scan_alternatives:range扫描分析
- index:索引名
- ranges:range扫描的条件范围
- index_dives_for_eq_ranges:是否使用了index dive,该值会被参数eq_range_index_dive_limit变量值影响。
- rowid_ordered:该range扫描的结果集是否根据PK值进行排序
- using_mrr:是否使用了mrr
- index_only:表示是否使用了覆盖索引
- rows:扫描的行数
- cost:索引的使用成本
- chosen:表示是否使用了该索引
- analyzing_roworder_intersect:分析是否使用了索引合并(index merge),如果未使用,会在cause中展示原因;如果使用了索引合并,会在该部分展示索引合并的代价。
- range_scan_alternatives:range扫描分析
-
chosen_range_access_summary:在前一个步骤中分析了各类索引使用的方法及代价,得出了一定的中间结果之后,在summary阶段汇总前一阶段的中间结果确认最后的方案
- range_access_plan:range扫描最终选择的执行计划。
- type:展示执行计划的type,如果使用了索引合并,则会显示index_roworder_intersect
- index:索引名
- rows:扫描的行数
- ranges:range扫描的条件范围
- rows_for_plan:该执行计划的扫描行数
- cost_for_plan:该执行计划的执行代价
- chosen:是否选择该执行计划
- range_access_plan:range扫描最终选择的执行计划。
-
considered_execution_plans
负责对比各可行计划的开销,并选择相对最优的执行计划。
{
"considered_execution_plans": [
{
"plan_prefix": [
] /* plan_prefix */,
"table": "`salaries`",
"best_access_path": {
"considered_access_paths": [
{
"access_type": "ref",
"index": "salaries_from_date_to_date_index",
"rows": 86,
"cost": 50.412,
"chosen": true
},
{
"access_type": "range",
"range_details": {
"used_index": "salaries_from_date_to_date_index"
} /* range_details */,
"chosen": false,
"cause": "heuristic_index_cheaper"
}
] /* considered_access_paths */
} /* best_access_path */,
"condition_filtering_pct": 100,
"rows_for_plan": 86,
"cost_for_plan": 50.412,
"chosen": true
}
] /* considered_execution_plans */
}
其中:
- plan_prefix:当前计划的前置执行计划。
- table:涉及的表名,如果有别名,也会展示出来
- best_access_path:通过对比considered_access_paths,选择一个最优的访问路径
- considered_access_paths:当前考虑的访问路径
- access_type:使用索引的方式,可参考explain中的type字段
- index:索引
- rows:行数
- cost:开销
- chosen:是否选用这种执行路径
- considered_access_paths:当前考虑的访问路径
- condition_filtering_pct:类似于explain的filtered列,是一个估算值
- rows_for_plan:执行计划最终的扫描行数,由considered_access_paths.rows X condition_filtering_pct计算获得。
- cost_for_plan:执行计划的代价,由considered_access_paths.cost相加获得
- chosen:是否选择了该执行计划
attaching_conditions_to_tables
基于considered_execution_plans中选择的执行计划,改造原有where条件,并针对表增加适当的附加条件,以便于单表数据的筛选。
TIPS
- 这部分条件的增加主要是为了便于ICP(索引条件下推),但ICP是否开启并不影响这部分内容的构造。
- ICP参考文档:https://www.cnblogs.com/Terry-Wu/p/9273177.html
{
"attaching_conditions_to_tables": {
"original_condition": "((`salaries`.`to_date` = DATE'1987-06-26') and (`salaries`.`from_date` = DATE'1986-06-26'))",
"attached_conditions_computation": [
] /* attached_conditions_computation */,
"attached_conditions_summary": [
{
"table": "`salaries`",
"attached": "((`salaries`.`to_date` = DATE'1987-06-26') and (`salaries`.`from_date` = DATE'1986-06-26'))"
}
] /* attached_conditions_summary */
} /* attaching_conditions_to_tables */
}
其中:
- original_condition:原始的条件语句
- attached_conditions_computation:使用启发式算法计算已使用的索引,如果已使用的索引的访问类型是ref,则计算用range能否使用组合索引中更多的列,如果可以,则用range的方式替换ref。
- attached_conditions_summary:附加之后的情况汇总
- table:表名
- attached:附加的条件或原语句中能直接下推给单表筛选的条件。
finalizing_table_conditions
最终的、经过优化后的表条件。
{
"finalizing_table_conditions": [
{
"table": "`salaries`",
"original_table_condition": "((`salaries`.`to_date` = DATE'1987-06-26') and (`salaries`.`from_date` = DATE'1986-06-26'))",
"final_table_condition ": null
}
] /* finalizing_table_conditions */
}
refine_plan
改善执行计划:
{
"refine_plan": [
{
"table": "`salaries`"
}
] /* refine_plan */
}
其中:
- table:表名及别名
join_execution
join_execution段落展示了执行阶段的执行过程。
"join_execution": {
"select#": 1,
"steps": [
] /* steps */
}
参考文档
相关文章
本文由博客一文多发平台 OpenWrite 发布!