cost 量化分析
作者：GreatSQL
- 2023-03-09 福建
本文字数：23662 字
阅读完需：约 78 分钟
GreatSQL 社区原创内容未经授权不得随意使用，转载请联系小编并注明来源。
GreatSQL 是 MySQL 的国产分支版本，使用上与 MySQL 一致。
作者： xryz
文章来源：GreatSQL 社区原创
前言：
我们在日常维护数据库的时候，经常会遇到查询慢的语句，这时候一般会通过执行 EXPLAIN 去查看它的执行计划，但是执行计划往往只给我们带来了最基础的分析信息，比如是否有使用索引，还有一些其他供我们分析的信息，比如使用了临时表、排序等等，却无法展示为什么一些其他的执行计划未被选择，比如说明明有索引，或者好几个索引，但是为什么查询时未使用到期望的索引等
explain select * from basic_person_info t1 join basic_person_info2 t2 on t1.id_num=t2.id_num where t1.age >10 and t2.age<20;+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+------+----------+-----------------------+| id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra |+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+------+----------+-----------------------+| 1 | SIMPLE | t2 | NULL | range | id_num_unique,idx_age,idx_age_id_num | idx_age | 1 | NULL | 9594 | 100.00 | Using index condition || 1 | SIMPLE | t1 | NULL | eq_ref | id_num_unique,idx_age | id_num_unique | 60 | test.t2.id_num | 1 | 50.00 | Using where |+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+------+----------+-----------------------+2 rows in set, 1 warning (0.01 sec)
复制代码
如上面这个例子，为什么 t2 表上列出了多个可能使用的索引，却选择了 idx_age，优化器为什么选择了指定的索引，这时候并不能直观的看出问题，这时候我们就可以开启 optimizer_trace 跟踪分析 MySQL 具体是怎么选择出最优的执行计划的。
OPTIMIZER_TRACE：
optimizer_trace 是什么：
optimizer_trace 是一个具有跟踪功能的工具，可以跟踪执行的语句的解析优化执行过程，并将跟踪到的信息记录到 INFORMATION_SCHEMA.OPTIMIZER_TRACE 表中，但是每个会话都只能跟踪它自己执行的语句，并且表中默认只记录最后一个查询的跟踪结果
使用方法：
# 打开optimizer trace功能 (默认情况下它是关闭的):set optimizer_trace="enabled=on";select ...; # 这里输入你自己的查询语句SELECT * FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE;# 当你停止查看语句的优化过程时，把optimizer trace功能关闭set optimizer_trace="enabled=off";
复制代码
相关参数：
mysql> show variables like '%optimizer_trace%';+------------------------------+----------------------------------------------------------------------------+| Variable_name | Value |+------------------------------+----------------------------------------------------------------------------+| optimizer_trace | enabled=on,one_line=off || optimizer_trace_features | greedy_search=on,range_optimizer=on,dynamic_range=on,repeated_subselect=on || optimizer_trace_limit | 1 || optimizer_trace_max_mem_size | 1048576 || optimizer_trace_offset | -1 |+------------------------------+----------------------------------------------------------------------------+
复制代码
optimizer_trace： enabled 开启/关闭 optimizer_trace，one_line 是否单行显示，关闭为 json 模式，一般不开启
optimizer_trace_features：跟踪信息中可打印的项，一般不调整默认打印所有项
optimizer_trace_limit：存储的跟踪 sql 条数
optimizer_trace_offset：开始记录的 sql 语句的偏移量，负数表示从最近执行倒数第几条开始记录
optimizer_trace_max_mem_size：optimizer_trace 内存的大小，如果跟踪信息超过这个大小，信息将会被截断
optimizer_trace 表信息：
该表总共有 4 个字段
QUERY 表示我们的查询语句。
TRACE 表示优化过程的 JSON 格式文本。(重点关注)
MISSING_BYTES_BEYOND_MAX_MEM_SIZE 由于优化过程可能会输出很多，如果超过某个限制时，多余的文本将不会被显示，这个字段展示了被忽略的文本字节数。
INSUFFICIENT_PRIVILEGES 表示是否没有权限查看优化过程，默认值是 0，只有某些特殊情况下才会是 1，我们暂时不关心这个字段的值。
信息解读：
通过 optimizer_trace 表的 query 字段可以看到，一条语句的执行过程主要分为三个步骤：
"join_preparation": {},（准备阶段）"join_optimization": {},（优化阶段）"join_execution": {},（执行阶段）
复制代码
各个步骤的详细内容解读：
preparation：
expanded_query ：将语句进行格式化，补充隐藏的列名和表名等transformations_to_nested_joins ：查询重写，比如join的on改为where语句
复制代码
optimization：
condition_processing{ ：条件句处理。 transformation{：转换类型句。这三次的转换分别是 equality_propagation（等值条件句转换），如：a = b and b = c and c = 5 constant_propagation（常量条件句转换），如：a = 1 AND b > a trivial_condition_removal（无效条件移除的转换），如：1 = 1 }}substitute_generated_columns ：替换虚拟生成列，测试了很多sql，这一列都没有看到有用的信息table_dependencies ：梳理表之间的依赖关系。ref_optimizer_key_uses ：如果优化器认为查询可以使用ref的话，在这里列出可以使用的索引。rows_estimation{ ：估算表行数和扫描的代价。如果查询中存在range扫描的话，对range扫描进行计划分析及代价估算。 table_scan：全表扫描的行数(rows)以及所需要的代价(cost)。 potential_range_indexes：该阶段会列出表中所有的索引并分析其是否可用，并且还会列出索引中可用的列字段。 analyzing_range_alternatives ：分析可选方案的代价。}considered_execution_plans{ ：对比各可行计划的代价，选择相对最优的执行计划。 plan_prefix：前置的执行计划。 best_access_path：当前最优的执行顺序信息结果集。 access_type表示使用索引的方式，可参照为explain中的type字段。 condition_filtering_pct：类似于explain中的filtered列，这是一个估算值。 rows_for_plan：该执行计划最终的扫描行数，这里的行数其实也是估算值，是由considered_access_paths的resulting_rows相乘之后再乘以condition_filtering_pct获得。 cost_for_plan：该执行计划的执行代价，由considered_access_paths的cost相加而得。 chosen：是否选择了该执行计划。}attaching_conditions_to_tables ：添加附加条件，使得条件尽可能筛选单表数据。refine_plan ：优化后的执行计划。
复制代码
trace 信息中的 json 信息很长，因为我们关心的是不同执行计划的 cost 区别，所以只需要重点关注两个部分 rows_estimation 和 considered_execution_plans
代价模型计算：
统计信息和 cost 计算参数：
计算 cost 会涉及到表的主键索引数据页(聚簇索引)数量和表中的记录数，两个信息都可以通过 innodb 的表统计信息 mysql.innodb_table_stats 查到，n_rows 是记录数，clustered_index_size 是聚簇索引页数。
mysql> select * from mysql.innodb_table_stats where table_name='basic_person_info';+---------------+-------------------+---------------------+--------+----------------------+--------------------------+| database_name | table_name | last_update | n_rows | clustered_index_size | sum_of_other_index_sizes |+---------------+-------------------+---------------------+--------+----------------------+--------------------------+| test | basic_person_info | 2022-12-23 18:27:24 | 86632 | 737 | 1401 |+---------------+-------------------+---------------------+--------+----------------------+--------------------------+1 row in set (0.01 sec)
复制代码
代价模型将操作分为 Server 层和 Engine（存储引擎）层两类，Server 层主要是 CPU 代价，Engine 层主要是 IO 代价，比如 MySQL 从磁盘读取一个数据页的代价 io_block_read_cost 为 1，从 buffer pool 读取的代价 memory_block_read_cost 为 0.25，计算符合条件的行代价为 row_evaluate_cost 为 0.1，除此之外还有：
memory_temptable_create_cost (default 1.0) 内存临时表的创建代价。
memory_temptable_row_cost (default 0.1) 内存临时表的行代价。
key_compare_cost (default 0.1) 键比较的代价，例如排序。
disk_temptable_create_cost (default 20.0) 内部 myisam 或 innodb 临时表的创建代价。
disk_temptable_row_cost (default 0.5) 内部 myisam 或 innodb 临时表的行代价。
这些都可以通过 mysql.server_cost、mysql.engine_cost 查看 defalt 值和设置值
mysql> select * from mysql.server_cost;+------------------------------+------------+---------------------+---------+---------------+| cost_name | cost_value | last_update | comment | default_value |+------------------------------+------------+---------------------+---------+---------------+| disk_temptable_create_cost | NULL | 2022-05-11 16:09:37 | NULL | 20 || disk_temptable_row_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.5 || key_compare_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.05 || memory_temptable_create_cost | NULL | 2022-05-11 16:09:37 | NULL | 1 || memory_temptable_row_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.1 || row_evaluate_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.1 |+------------------------------+------------+---------------------+---------+---------------+mysql> select * from mysql.engine_cost;+-------------+-------------+------------------------+------------+---------------------+---------+---------------+| engine_name | device_type | cost_name | cost_value | last_update | comment | default_value |+-------------+-------------+------------------------+------------+---------------------+---------+---------------+| default | 0 | io_block_read_cost | NULL | 2022-05-11 16:09:37 | NULL | 1 || default | 0 | memory_block_read_cost | NULL | 2023-01-09 11:17:39 | NULL | 0.25 |+-------------+-------------+------------------------+------------+---------------------+---------+---------------+
复制代码
计算公式：
如上面介绍的一样，代价模型将操作分为两类 io_cost 和 cpu_cost，io_cost+cpu_cost 就是总的 cost，下面是具体的计算方法：
全表扫描：
全表扫描成本 = io_cost + 1.1 + cpu_cost + 1 (io_cost +1.1 和 cpu_cost +1 在代码里是直接硬加上的，不知道为什么，计算的时候直接加上)
io_cost = clustered_index_size (统计信息中的主键页数) * avg_single_page_cost(读取一个页的平均成本)
avg_single_page_cost = pages_in_memory_percent * 0.25(memory_block_read_cost) + pages_on_disk_percent * 1.0(io_block_read_cost)
pages_in_memory_percent 表示已经加载到 Buffer Pool 中的叶结点占所有叶结点的比例 pages_on_disk_percent 表示没有加载到 Buffer Pool 中的叶结点占所有叶结点的比例
所以当数据已经全部读取到 buffer pool 中的时候：
io_cost=clustered_index_size * 0.25
都没有读取到 buffer pool 中的时候：
io_cost=clustered_index_size * 1.0
当部分数据在 buffer pool 中，部分数据需要从磁盘读取时，这时的系数介于 0.25 到 1 之间
cpu_cost = n_rows(统计信息中记录数) * 0.1(row_evaluate_cost)
走索引的成本：
和全表扫描的计算方法类似，其中 io_cost 与搜索的区间数有关，比如扫描三个区间 where a between 1 and 10 or a between 20 and 30 or a between 40 and 50，此时：
io_cost=3 * avg_single_page_cost
cpu_cost=记录数 * 0.1(row_evaluate_cost)+0.01(代码中的微调参数)
针对二级索引还会有回表的操作：
MySQL 认为每次回表都相当于是访问一个页面，所以每次回表都会进行一次 IO，这部分成本：
io_cost=rows(记录数)*avg_single_page_cost
对回表查询的数据还需要进行一次计算：
cpu_cost=rows(记录数) * 0.1(row_evaluate_cost)(需要注意的是当索引需要回表扫描时，在 rows_estimation 阶段并不会计算这个值，在 considered_execution_plans 阶段会重新加上这部分成本)
所以针对需要回表的查询：
io_cost=查询区间 * avg_single_page_cost + rows(记录数) * avg_single_page_cost
cpu_cost=记录数 * 0.1(row_evaluate_cost) + 0.01(代码中的微调参数) + rows(记录数) * 0.1(row_evaluate_cost)
例子：
mysql> set optimizer_trace='enabled=on';Query OK, 0 rows affected (0.00 sec)
mysql>explain select * from basic_person_info t1 join basic_person_info2 t2 on t1.id_num=t2.id_num where t1.age >10 and t2.age<20;+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+------+----------+-----------------------+| id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra |+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+------+----------+-----------------------+| 1 | SIMPLE | t2 | NULL | range | id_num_unique,idx_age,idx_age_id_num | idx_age | 1 | NULL | 9594 | 100.00 | Using index condition || 1 | SIMPLE | t1 | NULL | eq_ref | id_num_unique,idx_age | id_num_unique | 60 | test.t2.id_num | 1 | 50.00 | Using where |+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+------+----------+-----------------------+2 rows in set, 1 warning (0.04 sec)
复制代码
查看 optimizer_trace 的内容
select * from basic_person_info t1 join basic_person_info2 t2 on t1.id_num=t2.id_num where t1.age >10 and t2.age<20 | { "steps": [ { "join_preparation": { "select#": 1, "steps": [ { "expanded_query": "/* select#1 */ select `t1`.`id` AS `id`,`t1`.`id_num` AS `id_num`,`t1`.`lastname` AS `lastname`,`t1`.`firstname` AS `firstname`,`t1`.`mobile` AS `mobile`,`t1`.`sex` AS `sex`,`t1`.`birthday` AS `birthday`,`t1`.`age` AS `age`,`t1`.`top_education` AS `top_education`,`t1`.`address` AS `address`,`t1`.`income_by_year` AS `income_by_year`,`t1`.`create_time` AS `create_time`,`t1`.`update_time` AS `update_time`,`t2`.`id` AS `id`,`t2`.`id_num` AS `id_num`,`t2`.`lastname` AS `lastname`,`t2`.`firstname` AS `firstname`,`t2`.`mobile` AS `mobile`,`t2`.`sex` AS `sex`,`t2`.`birthday` AS `birthday`,`t2`.`age` AS `age`,`t2`.`top_education` AS `top_education`,`t2`.`address` AS `address`,`t2`.`income_by_year` AS `income_by_year`,`t2`.`create_time` AS `create_time`,`t2`.`update_time` AS `update_time` from (`basic_person_info` `t1` join `basic_person_info2` `t2` on((`t1`.`id_num` = `t2`.`id_num`))) where ((`t1`.`age` > 10) and (`t2`.`age` < 20))" }, { "transformations_to_nested_joins": { "transformations": [ "JOIN_condition_to_WHERE", "parenthesis_removal" ], "expanded_query": "/* select#1 */ select `t1`.`id` AS `id`,`t1`.`id_num` AS `id_num`,`t1`.`lastname` AS `lastname`,`t1`.`firstname` AS `firstname`,`t1`.`mobile` AS `mobile`,`t1`.`sex` AS `sex`,`t1`.`birthday` AS `birthday`,`t1`.`age` AS `age`,`t1`.`top_education` AS `top_education`,`t1`.`address` AS `address`,`t1`.`income_by_year` AS `income_by_year`,`t1`.`create_time` AS `create_time`,`t1`.`update_time` AS `update_time`,`t2`.`id` AS `id`,`t2`.`id_num` AS `id_num`,`t2`.`lastname` AS `lastname`,`t2`.`firstname` AS `firstname`,`t2`.`mobile` AS `mobile`,`t2`.`sex` AS `sex`,`t2`.`birthday` AS `birthday`,`t2`.`age` AS `age`,`t2`.`top_education` AS `top_education`,`t2`.`address` AS `address`,`t2`.`income_by_year` AS `income_by_year`,`t2`.`create_time` AS `create_time`,`t2`.`update_time` AS `update_time` from `basic_person_info` `t1` join `basic_person_info2` `t2` where ((`t1`.`age` > 10) and (`t2`.`age` < 20) and (`t1`.`id_num` = `t2`.`id_num`))" } } ] } }, { "join_optimization": { "select#": 1, "steps": [ { "condition_processing": { "condition": "WHERE", "original_condition": "((`t1`.`age` > 10) and (`t2`.`age` < 20) and (`t1`.`id_num` = `t2`.`id_num`))", "steps": [ { "transformation": "equality_propagation", "resulting_condition": "((`t1`.`age` > 10) and (`t2`.`age` < 20) and multiple equal(`t1`.`id_num`, `t2`.`id_num`))" }, { "transformation": "constant_propagation", "resulting_condition": "((`t1`.`age` > 10) and (`t2`.`age` < 20) and multiple equal(`t1`.`id_num`, `t2`.`id_num`))" }, { "transformation": "trivial_condition_removal", "resulting_condition": "((`t1`.`age` > 10) and (`t2`.`age` < 20) and multiple equal(`t1`.`id_num`, `t2`.`id_num`))" } ] } }, { "substitute_generated_columns": { } }, { "table_dependencies": [ { "table": "`basic_person_info` `t1`", "row_may_be_null": false, "map_bit": 0, "depends_on_map_bits": [ ] }, { "table": "`basic_person_info2` `t2`", "row_may_be_null": false, "map_bit": 1, "depends_on_map_bits": [ ] } ] }, { "ref_optimizer_key_uses": [ { "table": "`basic_person_info` `t1`", "field": "id_num", "equals": "`t2`.`id_num`", "null_rejecting": true }, { "table": "`basic_person_info2` `t2`", "field": "id_num", "equals": "`t1`.`id_num`", "null_rejecting": true } ] }, { "rows_estimation": [ { "table": "`basic_person_info` `t1`", "range_analysis": { "table_scan": { "rows": 86734, "cost": 8859.75 t1表的scan成本=聚簇索引页数*0.25 + 行数 * 0.1 +1.1+1 737*0.25+1.1+86734*0.1+1=8859.75 }, "potential_range_indexes": [ { "index": "PRIMARY", "usable": false, "cause": "not_applicable" }, { "index": "id_num_unique", "usable": false, "cause": "not_applicable" }, { "index": "mobile_unique", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_name", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_top_education", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_create_time", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_mobile", "usable": false, "cause": "not_applicable" }, { "index": "idx_age", "usable": true, "key_parts": [ "age", "id" ] } ], "setup_range_conditions": [ ], "group_index_range": { "chosen": false, "cause": "not_single_table" }, "skip_scan_range": { "chosen": false, "cause": "not_single_table" }, "analyzing_range_alternatives": { "range_scan_alternatives": [ { "index": "idx_age", "ranges": [ "10 < age" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": false, "rows": 43367, "cost": 15178.7, 通过索引idx_age读取数据： io_cost=区间数* 0.25 +记录数* 0.25 io_cost=1*0.25+43367*0.25=10,842 cpu_cost=记录数* 0.1 (没有回表的cost) cpu_cost=43367*0.1=4,336.7 cost=10842+4,336.7=15178.7 "chosen": false, "cause": "cost" } ], "analyzing_roworder_intersect": { "usable": false, "cause": "too_few_roworder_scans" } } } }, { "table": "`basic_person_info2` `t2`", "range_analysis": { "table_scan": { "rows": 73845, "cost": 7538.85 t2表的scan成本=聚簇索引页数*0.25 + 行数 * 0.1 +1.1+1 609*0.25+1+73845*0.1+1.1=7538.85 }, "potential_range_indexes": [ { "index": "PRIMARY", "usable": false, "cause": "not_applicable" }, { "index": "id_num_unique", "usable": false, "cause": "not_applicable" }, { "index": "mobile_unique", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_name", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_top_education", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_create_time", "usable": false, "cause": "not_applicable" }, { "index": "idx_basic_person_info_mobile", "usable": false, "cause": "not_applicable" }, { "index": "idx_age", "usable": true, "key_parts": [ "age", "id" ] }, { "index": "idx_age_id_num", "usable": true, "key_parts": [ "age", "id_num", "id" ] } ], "setup_range_conditions": [ ], "group_index_range": { "chosen": false, "cause": "not_single_table" }, "skip_scan_range": { "chosen": false, "cause": "not_single_table" }, "analyzing_range_alternatives": { "range_scan_alternatives": [ { "index": "idx_age", "ranges": [ "age < 20" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": false, "rows": 9594, "cost": 3358.16, 通过索引idx_age读取数据： io_cost=区间数* 0.25 +记录数* 0.25 io_cost=1*0.25+9594*0.25=2,398.75 cpu_cost=记录数* 0.1 (没有回表的cost) cpu_cost=9594*0.1959.4 cost=2,398.75+959.4=3,358.15 "chosen": true }, { "index": "idx_age_id_num", "ranges": [ "age < 20" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": false, "rows": 19086, "cost": 6680.36, 通过索引idx_age_id_num读取数据： io_cost=区间数* 0.25 +记录数* 0.25 io_cost=1*0.25+19086*0.25=4,771.75 cpu_cost=记录数* 0.1 (没有回表的cost) cpu_cost=19086*0.1=1908.6 cost=4,771.75+1908.6=6,680.35 "chosen": false, "cause": "cost" } ], "analyzing_roworder_intersect": { "usable": false, "cause": "too_few_roworder_scans" } }, "chosen_range_access_summary": { "range_access_plan": { "type": "range_scan", "index": "idx_age", "rows": 9594, "ranges": [ "age < 20" ] }, "rows_for_plan": 9594, "cost_for_plan": 3358.16, "chosen": true } } } ] }, { "considered_execution_plans": [ { "plan_prefix": [ ], "table": "`basic_person_info2` `t2`", "best_access_path": { "considered_access_paths": [ { "access_type": "ref", "index": "id_num_unique", "usable": false, "chosen": false }, { "rows_to_scan": 9594, "filtering_effect": [ ], "final_filtering_effect": 1, "access_type": "range", "range_details": { "used_index": "idx_age" }, "resulting_rows": 9594, "cost": 4317.56, 通过索引idx_age读取数据： io_cost=区间数* 0.25 +记录数* 0.25 io_cost=1*0.25+9594*0.25=2,398.75 cpu_cost=记录数* 0.1 + 记录数* 0.1 cpu_cost=9594*0.1*2=1,918.8 cost=2,398.75+1,918.8=4317.56 "chosen": true } ] }, "condition_filtering_pct": 100, "rows_for_plan": 9594, "cost_for_plan": 4317.56, "rest_of_plan": [ { "plan_prefix": [ "`basic_person_info2` `t2`" ], "table": "`basic_person_info` `t1`", "best_access_path": { "considered_access_paths": [ { "access_type": "eq_ref", "index": "id_num_unique", "rows": 1, "cost": 3357.9, io_cost=t2表记录数*0.25=9594*0.25=2398.5 cpu_cost=记录数*0.1=9594*0.1=959.4 cost=2398.5+959.4=3357.9 "chosen": true }, { "rows_to_scan": 86734, "filtering_effect": [ ], "final_filtering_effect": 0.5, "access_type": "scan", "using_join_cache": true, "buffers_needed": 14, "resulting_rows": 43367, "cost": 4.16701e+07, "chosen": false } ] }, "condition_filtering_pct": 100, "rows_for_plan": 9594, "cost_for_plan": 7675.46, 总cost=4,317.56+3,357.9=7,675.46 "chosen": true } ] }, { "plan_prefix": [ ], "table": "`basic_person_info` `t1`", "best_access_path": { "considered_access_paths": [ { "access_type": "ref", "index": "id_num_unique", "usable": false, "chosen": false }, { "rows_to_scan": 86734, "filtering_effect": [ ], "final_filtering_effect": 0.5, "access_type": "scan", "resulting_rows": 43367, "cost": 8857.65, t1的scan成本 "chosen": true } ] }, "condition_filtering_pct": 100, "rows_for_plan": 43367, "cost_for_plan": 8857.65, "pruned_by_cost": true 放弃后续的计算 } ] }, { "attaching_conditions_to_tables": { "original_condition": "((`t1`.`id_num` = `t2`.`id_num`) and (`t1`.`age` > 10) and (`t2`.`age` < 20))", "attached_conditions_computation": [ ], "attached_conditions_summary": [ { "table": "`basic_person_info2` `t2`", "attached": "(`t2`.`age` < 20)" }, { "table": "`basic_person_info` `t1`", "attached": "((`t1`.`id_num` = `t2`.`id_num`) and (`t1`.`age` > 10))" } ] } }, { "finalizing_table_conditions": [ { "table": "`basic_person_info2` `t2`", "original_table_condition": "(`t2`.`age` < 20)", "final_table_condition ": "(`t2`.`age` < 20)" }, { "table": "`basic_person_info` `t1`", "original_table_condition": "((`t1`.`id_num` = `t2`.`id_num`) and (`t1`.`age` > 10))", "final_table_condition ": "(`t1`.`age` > 10)" } ] }, { "refine_plan": [ { "table": "`basic_person_info2` `t2`", "pushed_index_condition": "(`t2`.`age` < 20)", "table_condition_attached": null }, { "table": "`basic_person_info` `t1`" } ] } ] } }, { "join_execution": { "select#": 1, "steps": [ ] } } ]}
复制代码
成本常数修改：
前面已经介绍了成本常量值实际上存放在 MySQL 自带的系统库 MySQL 中的 server_cost 和 engine_cost 表中，其中 server_cost 表存放 server 层的成本常量，engine_cost 表存放 engine 层成本常量
mysql> select * from mysql.server_cost;+------------------------------+------------+---------------------+---------+---------------+| cost_name | cost_value | last_update | comment | default_value |+------------------------------+------------+---------------------+---------+---------------+| disk_temptable_create_cost | NULL | 2022-05-11 16:09:37 | NULL | 20 || disk_temptable_row_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.5 || key_compare_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.05 || memory_temptable_create_cost | NULL | 2022-05-11 16:09:37 | NULL | 1 || memory_temptable_row_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.1 || row_evaluate_cost | NULL | 2022-05-11 16:09:37 | NULL | 0.1 |+------------------------------+------------+---------------------+---------+---------------+
mysql> select * from mysql.engine_cost;+-------------+-------------+------------------------+------------+---------------------+---------+---------------+| engine_name | device_type | cost_name | cost_value | last_update | comment | default_value |+-------------+-------------+------------------------+------------+---------------------+---------+---------------+| default | 0 | io_block_read_cost | NULL | 2022-05-11 16:09:37 | NULL | 1 || default | 0 | memory_block_read_cost | NULL | 2023-01-09 11:17:39 | NULL | 0.25 |+-------------+-------------+------------------------+------------+---------------------+---------+---------------+
复制代码
其中 default_value 的值是系统默认的，不能修改，cost_value 列的值我们可以修改，如果 cost_value 列的值不为空系统将用该值覆盖默认值，我们可以通过 update 语句来修改
mysql> update mysql.engine_cost set cost_value=10 where cost_name='memory_block_read_cost';Query OK, 0 rows affected (0.00 sec)mysql> update mysql.engine_cost set cost_value=10 where cost_name='io_block_read_cost';Query OK, 0 rows affected (0.00 sec)
复制代码
很多资料都说执行 flush optimizer_costs 就可以生效，不过我在修改完后并执行 flush optimizer_costs 并不能马上生效，最后是通过重启数据库实例才生效，这个可能是数据库版本的差异，大家可以自行验证。
mysql> explain select * from basic_person_info t1 join basic_person_info2 t2 on t1.id_num=t2.id_num where t1.age >10 and t2.age<20;+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+-------+----------+-------------+| id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra |+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+-------+----------+-------------+| 1 | SIMPLE | t2 | NULL | ALL | id_num_unique,idx_age,idx_age_id_num | NULL | NULL | NULL | 73990 | 12.97 | Using where || 1 | SIMPLE | t1 | NULL | eq_ref | id_num_unique,idx_age | id_num_unique | 60 | test.t2.id_num | 1 | 50.00 | Using where |+----+-------------+-------+------------+--------+--------------------------------------+---------------+---------+----------------+-------+----------+-------------+
"table": "`basic_person_info2` `t2`", "range_analysis": { "table_scan": { "rows": 73990, "cost": 13491.1 全表扫描cost=609*10+73990*0.1+1.1+1= 13491.1 },"index": "idx_age", "ranges": [ "age < 20" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": false, "rows": 9594, "cost": 96909.4, idx_age索引扫描cost=1*10+9594*10+9594*0.1=96,909.4 "chosen": false, "cause": "cost" },
复制代码
修改后的执行计划，发现 t2 表走了全表扫描了而没有走 idx_age 索引，分别查看一下 t2 表走全表扫描和 idx_age 索引的 cost 发现全表扫描的 cost 为 13491.1，而走索引的 cost 为 96,909.4，因为全表扫描的 cost 比走索引低，所以优化器没有选择 idx_age 索引。
从这个例子可以看出，更改成本常量值会直接影响优化器的方案选择，所以一定要慎重，没有特殊原因建议不要修改。
explain format=json
虽然通过 optimizer_trace 可以看到很多详细的优化器选择过程，但是使用起来起来还是比较麻烦，需要过滤的信息很多，这时 explain format=json 输出 json 格式的分析数据也是一个不错的选择，它也包含语句将要执行的成本信息，如下：
query_cost 总查询成本read_cost IO成本+除 eval_cost以外cpu成本eval_cost 检测rows * filter条记录的成本prefix_cost 单次查询的成本,等于read_cost+eval_costmysql> explain format=json select * from basic_person_info t1 join basic_person_info2 t2 on t1.id_num=t2.id_num where t1.age >10 and t2.age<20;{ "query_block": { "select_id": 1, "cost_info": { "query_cost": "7675.46" }, "nested_loop": [ { "table": { "table_name": "t2", "access_type": "range", "possible_keys": [ "id_num_unique", "idx_age", "idx_age_id_num" ], "key": "idx_age", "used_key_parts": [ "age" ], "key_length": "1", "rows_examined_per_scan": 9594, "rows_produced_per_join": 9594, "filtered": "100.00", "index_condition": "(`test`.`t2`.`age` < 20)", "cost_info": { "read_cost": "3358.16", 包含所有io成本+(cpu成本-eval_cost) "eval_cost": "959.40", 计算扇出的cpu成本，优化器利用启发式规则估算出满足所有条件的的比例(filtered) =rows_examined_per_scan*filtered*0.1 "prefix_cost": "4317.56", 单表查询的总成本 "data_read_per_join": "3M" }, "used_columns": [ "id", "id_num", "lastname", "firstname", "mobile", "sex", "birthday", "age", "top_education", "address", "income_by_year", "create_time", "update_time" ] } }, { "table": { "table_name": "t1", "access_type": "eq_ref", "possible_keys": [ "id_num_unique", "idx_age" ], "key": "id_num_unique", "used_key_parts": [ "id_num" ], "key_length": "60", "ref": [ "test.t2.id_num" ], "rows_examined_per_scan": 1, "rows_produced_per_join": 4797, "filtered": "50.00", "cost_info": { "read_cost": "2398.50", 包含所有io成本+(cpu成本-eval_cost) "eval_cost": "479.70", 计算扇出的cpu成本，优化器利用启发式规则估算出满足所有条件的的比例(filtered) =rows_examined_per_scan*filtered*0.1 "prefix_cost": "7675.46", 两表查询的总cost "data_read_per_join": "1M" }, "used_columns": [ "id", "id_num", "lastname", "firstname", "mobile", "sex", "birthday", "age", "top_education", "address", "income_by_year", "create_time", "update_time" ], "attached_condition": "(`test`.`t1`.`age` > 10)" } } ] }}
复制代码
另外，explain 结合 show warnings 语句一起使用还可以得知优化器改写后的语句
mysql> show warnings;+-------+------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+| Level | Code | Message |+-------+------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+| Note | 1003 | /* select#1 */ select `test`.`t1`.`id` AS `id`,`test`.`t1`.`id_num` AS `id_num`,`test`.`t1`.`lastname` AS `lastname`,`test`.`t1`.`firstname` AS `firstname`,`test`.`t1`.`mobile` AS `mobile`,`test`.`t1`.`sex` AS `sex`,`test`.`t1`.`birthday` AS `birthday`,`test`.`t1`.`age` AS `age`,`test`.`t1`.`top_education` AS `top_education`,`test`.`t1`.`address` AS `address`,`test`.`t1`.`income_by_year` AS `income_by_year`,`test`.`t1`.`create_time` AS `create_time`,`test`.`t1`.`update_time` AS `update_time`,`test`.`t2`.`id` AS `id`,`test`.`t2`.`id_num` AS `id_num`,`test`.`t2`.`lastname` AS `lastname`,`test`.`t2`.`firstname` AS `firstname`,`test`.`t2`.`mobile` AS `mobile`,`test`.`t2`.`sex` AS `sex`,`test`.`t2`.`birthday` AS `birthday`,`test`.`t2`.`age` AS `age`,`test`.`t2`.`top_education` AS `top_education`,`test`.`t2`.`address` AS `address`,`test`.`t2`.`income_by_year` AS `income_by_year`,`test`.`t2`.`create_time` AS `create_time`,`test`.`t2`.`update_time` AS `update_time` from `test`.`basic_person_info` `t1` join `test`.`basic_person_info2` `t2` where ((`test`.`t1`.`id_num` = `test`.`t2`.`id_num`) and (`test`.`t1`.`age` > 10) and (`test`.`t2`.`age` < 20)) |+-------+------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+1 row in set (0.00 sec)
复制代码
总结：
MySQL 的优化器是基于成本来选择最优执行方案的，哪个成本最少就选哪个，所以重点在于计算出各个执行计划的 cost
成本由 CPU 成本和 IO 成本组成，每个成本常数值可以自己调整，非必要的情况下不要调整，以免影响整个数据库的执行计划选择
通过开启 optimizer_trace 可以跟踪优化器的各个环节的分析步骤，可以判断有时候为什么没有走索引而走了全表扫描
explain 加上 format=json 选项后可以查看成本信息分为 read_cost 和 eval_cost，但只能看到当前已经选择的执行计划，另外通过 show warnings 可以看到优化器改写后的语句
Enjoy GreatSQL :)
关于 GreatSQL
GreatSQL 是由万里数据库维护的 MySQL 分支，专注于提升 MGR 可靠性及性能，支持 InnoDB 并行查询特性，是适用于金融级应用的 MySQL 分支版本。
相关链接： GreatSQL社区 Gitee GitHub Bilibili
GreatSQL 社区：
社区博客有奖征稿详情：https://greatsql.cn/thread-100-1-1.html
技术交流群：
微信：扫码添加
GreatSQL社区助手微信好友，发送验证信息加群。
划线
评论
复制
发布于: 刚刚阅读数: 3
版权声明: 本文为 InfoQ 作者【GreatSQL】的原创文章。
原文链接:【http://xie.infoq.cn/article/5df3a126ae5277612f90a92a0】。文章转载请联系作者。
GreatSQL
关注
GreatSQL社区 2023-01-31 加入
GreatSQL是由万里数据库维护的MySQL分支，专注于提升MGR可靠性及性能，支持InnoDB并行查询特性，是适用于金融级应用的MySQL分支版本。 社区：https://greatsql.cn/ Gitee: https://gitee.com/GreatSQL/GreatSQL
评论