技术分享 | 为什么 SELECT 查询选择全表扫描，而不走索引？

1、基本结论

SQL 的执行成本（cost）是 MySQL 优化器选择 SQL 执行计划时一个重要考量因素。当优化器认为使用索引的成本高于全表扫描的时候，优化器将会选择全表扫描，而不是使用索引。

下面通过一个实验来说明。

2、问题现象

如下结构的一张表，表中约有104w行数据：

CREATE TABLE `test03` (
  `id` int(11) NOT NULL AUTO_INCREMENT COMMENT '自增主键',
  `dept` tinyint(4) NOT NULL COMMENT '部门id',
  `name` varchar(30) COLLATE utf8mb4_bin DEFAULT NULL COMMENT '用户名称',
  `create_time` datetime NOT NULL COMMENT '注册时间',
  `last_login_time` datetime DEFAULT NULL COMMENT '最后登录时间',
  PRIMARY KEY (`id`),
  KEY `ct_index` (`create_time`)
) ENGINE=InnoDB AUTO_INCREMENT=1048577 DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin COMMENT='测试表'

• type 为 ALL ，而不是 range
• rows 行数和全表行数接近

# 查询1
mysql> explain select * from test03 where create_time > '2021-10-01 02:04:36';
+----+-------------+--------+------------+------+---------------+------+---------+------+---------+----------+-------------+
| id | select_type | table  | partitions | type | possible_keys | key  | key_len | ref  | rows    | filtered | Extra       |
+----+-------------+--------+------------+------+---------------+------+---------+------+---------+----------+-------------+
|  1 | SIMPLE      | test03 | NULL       | ALL  | ct_index      | NULL | NULL    | NULL | 1045955 |    50.00 | Using where |
+----+-------------+--------+------------+------+---------------+------+---------+------+---------+----------+-------------+
1 row in set, 1 warning (0.00 sec)

而查询2，则用到了 ct_index(create_time) 索引:

# 查询2
mysql> explain select * from test03 where create_time < '2021-01-01 02:04:36';
+----+-------------+--------+------------+-------+---------------+----------+---------+------+------+----------+-----------------------+
| id | select_type | table  | partitions | type  | possible_keys | key      | key_len | ref  | rows | filtered | Extra                 |
+----+-------------+--------+------------+-------+---------------+----------+---------+------+------+----------+-----------------------+
|  1 | SIMPLE      | test03 | NULL       | range | ct_index      | ct_index | 5       | NULL |  169 |   100.00 | Using index condition |
+----+-------------+--------+------------+-------+---------------+----------+---------+------+------+----------+-----------------------+

3、获得SQL优化器处理信息

这里使用 optimizer trace 工具，观察 MySQL 对 SQL 的优化处理过程：

# 调大trace的容量，防止被截断
set global optimizer_trace_max_mem_size = 1048576;

# 开启optimizer_trace
set optimizer_trace="enabled=on";

# 执行SQL
select * from test03 where create_time > '2021-10-01 02:04:36';

# SQL执行完成之后，查看TRACE
 select TRACE from INFORMATION_SCHEMA.OPTIMIZER_TRACE\G

获得关于此 SQL 的详细优化器处理信息：

mysql> select TRACE from INFORMATION_SCHEMA.OPTIMIZER_TRACE\G                                                                                       
*************************** 1. row ***************************
TRACE: {
  "steps": [
    {
      "join_preparation": {
        "select#": 1,
        "steps": [
          {
            "expanded_query": "/* select#1 */ select `test03`.`id` AS `id`,`test03`.`dept` AS `dept`,`test03`.`name` AS `name`,`test03`.`create_time` AS `create_time`,`test03`.`last_login_time` AS `last_login_time` from `test03` where (`test03`.`create_time` > '2021-10-01 02:04:36')"
          }
        ]
      }
    },
    {
      "join_optimization": {
        "select#": 1,
        "steps": [
          {
            "condition_processing": {
              "condition": "WHERE",
              "original_condition": "(`test03`.`create_time` > '2021-10-01 02:04:36')",
              "steps": [
                {
                  "transformation": "equality_propagation",
                  "resulting_condition": "(`test03`.`create_time` > '2021-10-01 02:04:36')"
                },
                {
                  "transformation": "constant_propagation",
                  "resulting_condition": "(`test03`.`create_time` > '2021-10-01 02:04:36')"
                },
                {
                  "transformation": "trivial_condition_removal",
                  "resulting_condition": "(`test03`.`create_time` > '2021-10-01 02:04:36')"
                }
              ]
            }
          },
          {
            "substitute_generated_columns": {
            }
          },
          {
            "table_dependencies": [
              {
                "table": "`test03`",
                "row_may_be_null": false,
                "map_bit": 0,
                "depends_on_map_bits": [
                ]
              }
            ]
          },
          {
            "ref_optimizer_key_uses": [
            ]
          },
          {
            "rows_estimation": [
              {
                "table": "`test03`",
                "range_analysis": {
                  "table_scan": {
                    "rows": 1045955,
                    "cost": 212430
                  },
                  "potential_range_indexes": [
                    {
                      "index": "PRIMARY",
                      "usable": false,
                      "cause": "not_applicable"
                    },
                    {
                      "index": "ct_index",
                      "usable": true,
                      "key_parts": [
                        "create_time",
                        "id"
                      ]
                    }
                  ],
                  "setup_range_conditions": [
                  ],
                  "group_index_range": {
                    "chosen": false,
                    "cause": "not_group_by_or_distinct"
                  },
                  "analyzing_range_alternatives": {
                    "range_scan_alternatives": [
                      {
                        "index": "ct_index",
                        "ranges": [
                          "0x99aac22124 < create_time"
                        ],
                        "index_dives_for_eq_ranges": true,
                        "rowid_ordered": false,
                        "using_mrr": false,
                        "index_only": false,
                        "rows": 522977,
                        "cost": 627573,
                        "chosen": false,
                        "cause": "cost"
                      }
                    ],
                    "analyzing_roworder_intersect": {
                      "usable": false,
                      "cause": "too_few_roworder_scans"
                    }
                  }
                }
              }
            ]
          },
          {
            "considered_execution_plans": [
              {
                "plan_prefix": [
                ],
                "table": "`test03`",
                "best_access_path": {
                  "considered_access_paths": [
                    {
                      "rows_to_scan": 1045955,
                      "access_type": "scan",
                      "resulting_rows": 1.05e6,
                      "cost": 212428,
                      "chosen": true
                    }
                  ]
                },
                "condition_filtering_pct": 100,
                "rows_for_plan": 1.05e6,
                "cost_for_plan": 212428,
                "chosen": true
              }
            ]
          },
          {
            "attaching_conditions_to_tables": {
              "original_condition": "(`test03`.`create_time` > '2021-10-01 02:04:36')",
              "attached_conditions_computation": [
              ],
              "attached_conditions_summary": [
                {
                  "table": "`test03`",
                  "attached": "(`test03`.`create_time` > '2021-10-01 02:04:36')"
                }
              ]
            }
          },
          {
            "refine_plan": [
              {
                "table": "`test03`"
              }
            ]
          }
        ]
      }
    },
    {
      "join_execution": {
        "select#": 1,
        "steps": [
        ]
      }
    }
  ]
}
1 row in set (0.00 sec)

• 明确指出了使用索引 ct_index(create_time) 和全表扫描的成本差异

• 同时指出了未选择索引的原因：cost

4、为什么使用索引的成本比全表扫描还高？

通过观察优化器的信息，不难发现，使用索引扫描行数约52w行，而全表扫描约为104w行。为什么优化器反而认为使用索引的成本比全表扫描还高呢？

因为当 ct_index(create_time) 这个普通索引并不包括查询的所有列，因此需要通过 ct_index 的索引树找到对应的主键 id ，然后再到 id 的索引树进行数据查询，即回表（通过索引查出主键，再去查数据行），这样成本必然上升。尤其是当回表的数据量比较大的时候，经常会出现 MySQL 优化器认为回表查询代价过高而不选择索引的情况。

这里可以回头看查询1和查询2的数据量占比：

• 查询1的数据量占整个表的60%，回表成本高，因此优化器选择了全表扫描

• 查询2的数据量占整个表的0.02%，因此优化器选择了索引

mysql> select (select count(*) from test03 where create_time > '2021-10-01 02:04:36')/(select count(*) from test03) as '>20211001', (select count(*) from test03 where create_time < '2021-01-01 02:04:36')/(select count(*) from test03) as '<20210101';
+-----------+-----------+
| >20211001 | <20210101 |
+-----------+-----------+
|    0.5997 |    0.0002 |
+-----------+-----------+
1 row in set (0.44 sec)

• 当优化器认为全表扫描成本更低的时候，就不会使用索引

• 并没有一个固定的数据量占比来决定优化器是否使用全表扫描（曾经是30%）

• 优化器在选择的时候会考虑更多的因素，如：表大小，行数量，IO块大小等

https://dev.mysql.com/doc/refman/5.7/en/where-optimization.html