【GreatSQL优化器-05】条件过滤condition_fanout_filter
【GreatSQL优化器-05】条件过滤condition_fanout_filter一、condition_fanout_filter介绍
GreatSQL 的优化器对于 join 的表需要根据行数和 cost 来确定最后哪张表先执行哪张表后执行，这里面就涉及到预估满足条件的表数据，condition_fanout_filter 会根据一系列方法计算出一个数据过滤百分比，这个比百分比就是 filtered 系数，这个值区间在，值越小代表过滤效果越好。用这个系数乘以总的行数就可以得出最后需要扫描的表行数的数量，可以大幅节省开销和执行时间。
这个功能是由 OPTIMIZER_SWITCH_COND_FANOUT_FILTER这个OPTIMIZER_SWITCH 来控制的，默认是打开的。因此一般情况下不需要特意去关闭，但是如果遇到执行特别慢的一些情况可以考虑关闭。
下面用一个简单的例子来说明 condition_fanout_filter 是什么：
CREATE TABLE t1 (c1 INT PRIMARY KEY, c2 INT,date1 DATETIME);
INSERT INTO t1 VALUES (1,10,'2021-03-25 16:44:00.123456'),(2,1,'2022-03-26 16:44:00.123456'),(3,4,'2023-03-27 16:44:00.123456'),(5,5,'2024-03-25 16:44:00.123456');
CREATE TABLE t2 (cc1 INT PRIMARY KEY, cc2 INT);
INSERT INTO t2 VALUES (1,3),(2,1),(3,2),(4,3),(5,15);
# 为了查看过滤系数，需要创建一张没有主键的表用来做过滤估计。
CREATE TABLE t3 (ccc1 INT, ccc2 varchar(100));
INSERT INTO t3 VALUES (1,'aa1'),(2,'bb1'),(3,'cc1'),(4,'dd1'),(null,'ee');
CREATE INDEX idx1 ON t1(c2);
CREATE INDEX idx2 ON t1(c2,date1);
CREATE INDEX idx2_1 ON t2(cc2);
CREATE INDEX idx3_1 ON t3(ccc1);看到下面的 t3 的过滤百分比46.66%，意味着两张表 join 一共20行，因为 t3 的过滤百分比为 46.66%，因此最后只需要读取 20*46.66%=9 行数据。
注意，这里没有建立直方图，因此结果不包含直方图过滤的因素，关于直方图后面会专门开一章讲。
greatsql> EXPLAIN SELECT * FROM t1 JOIN t3 ON t1.c1=t3.ccc1 OR t3.ccc1 <3;
+----+-------------+-------+------------+-------+---------------+------+---------+------+------+----------+--------------------------------------------+
| id | select_type | table | partitions | type| possible_keys | key| key_len | ref| rows | filtered | Extra |
+----+-------------+-------+------------+-------+---------------+------+---------+------+------+----------+--------------------------------------------+
|1 | SIMPLE | t1 | NULL | index | PRIMARY | idx2 | 11 | NULL | 4 | 100.00 | Using index |
|1 | SIMPLE | t3 | NULL | ALL | idx3_1 | NULL | NULL | NULL | 5 | 46.66 | Using where; Using join buffer (hash join) |
# 这里显示的值就是 t3 的过滤数据百分比
+----+-------------+-------+------------+-------+---------------+------+---------+------+------+----------+--------------------------------------------+
greatsql> EXPLAIN FORMAT=TREE SELECT * FROM t1 JOIN t3 ON t1.c1=t3.ccc1 OR t3.ccc1<3;
+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| EXPLAIN |
+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| -> Filter: ((t3.ccc1 = t1.c1) or (t3.ccc1 < 3))(cost=3.65 rows=9)
-> Inner hash join (no condition)(cost=3.65 rows=9) # 这里结果为读取9行数据，跟上面算出来的数据一致
-> Table scan on t3(cost=0.12 rows=5)
-> Hash
-> Index scan on t1 using idx2(cost=1.40 rows=4)
|
+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+二、best_access_path代码解释
condition_fanout_filte的计算在 best_access_path函数实现，用来预估不同 join 连接时候的 join 表的读取行数和可能的 cost。
void Optimize_table_order::best_access_path(JOIN_TAB *tab,
const table_map remaining_tables,
const uint idx, bool disable_jbuf,
const double prefix_rowcount,
POSITION *pos) {
# 如果根据前面的结果keyuse_array数组有值的话，那么根据find_best_ref()函数先找出最优索引，按照索引的方式计算cost
if (tab->keyuse() != nullptr &&
(table->file->ha_table_flags() & HA_NO_INDEX_ACCESS) == 0)
best_ref =
find_best_ref(tab, remaining_tables, idx, prefix_rowcount,
&found_condition, &ref_depend_map, &used_key_parts);
# 最主要计算下面3个值
pos->filter_effect = filter_effect = std::(1.0, tab->found_records * calculate_condition_filter() / rows_after_filtering);
pos->rows_fetched = rows_fetched = rows_after_filtering;
pos->read_cost = scan_read_cost = calculate_scan_cost();
}下面是代码里面涉及的计算公式，这里是 keyuse_array 数组为空的情况，也就是扫描方式 "access_type" 非 "eq_ref" 和 "ref" 的情况，或者说是没有找到最优索引的情况。keyuse_array 数组有值的情况，在函数find_best_ref()计算，结果有可能也走下面的计算方式，详细在后面的章节详细介绍。
关键参数解释值rows_fetched总共需要读取多少行rows_after_filtering 见下表一filter_effect条件过滤百分比系数std::min(1.0, tab->found_records * calculate_condition_filter() / rows_after_filtering) 见下表一和二read_cost读的开销calculate_scan_cost() 见下表四prefix_rowcountjoin左表的行数，意味着多少行会被join到右表 对于第一张表来说prefix_rowcount=1第一张表：prefix_rowcount = rows_fetched * filter_effect 非第一张表：prefix_rowcount = 上一张表的prefix_rowcount * rows_fetched * filter_effectprefix_costjoin左表的cost，row_evaluate_cost()计算公式=0.1 * 行数，0.1是读一行的开销第一张表：read_cost + cm->row_evaluate_cost(prefix_rowcount) 非第一张表：上一张表的prefix_cost + read_cost + cm->row_evaluate_cost(上一张表的prefix_rowcount*rows_fetched)表一，rows_after_filtering 计算方式
场景解释值OPTIMIZER_SWITCH_COND_FANOUT_FILTER模式(默认ON)条件过滤模式开启tab->found_records * calculate_condition_filter() 见下表二table->quick_condition_rows != tab->found_records通过别的方法获取的表满足的行数table->quick_condition_rowskeyuse_array有值(参考前面)REF扫描模式tab->found_records * 0.75以上都不符合。默认情况全表行数tab->found_records表二，calculate_condition_filter() 计算方式
场景值说明满足条件的表的行数为0filter = COND_FILTER_ALLPASS 见表三使用索引filter = filter * std::min(table->quick_rows / tab->records() , 1.0)带有条件并且条件里涉及不含有索引的列filter = filter * Item_cond->get_filtering_effect() 见表三合并结果filter = max(filter, 1.0 / tab->records())(filter * fanout) < 0.05filter = 0.05 / fanoutfanout是满足条件的表的行数注：filter是条件过滤百分比，这个值区间在，这个值越小代表过滤效果越好
表三，get_filtering_effect() 算出来的过滤系数
Item的系数解释系数说明COND_FILTER_ALLPASSAlways true1.0f代表全部数据都符合，全部都要扫描COND_FILTER_EQUALITYcol1 = col20.1f代表预估10%数据符合条件COND_FILTER_INEQUALITYcol1 > col20.3333f代表预估1/3数据符合条件COND_FILTER_BETWEENcol1 BETWEEN a AND b0.1111f代表预估1/9数据符合条件表四，calculate_scan_cost() 计算方式
场景值说明如果是范围扫描prefix_rowcount * (tab->range_scan()->cost + cost_model->row_evaluate_cost(tab->found_records - *rows_after_filtering))非join buffer模式prefix_rowcount * (single_scan_read_cost + cost_model->row_evaluate_cost(tab->records() - *rows_after_filtering)) single_scan_read_cost计算如下： force_index模式: table->file->read_cost(tab->ref().key, 1,tab->records() 见表五) 非force_index模式: table->file->table_scan_cost() 见表五join buffer模式(默认ON)buffer_count * (single_scan_read_cost+ cost_model->row_evaluate_cost(tab->records() - *rows_after_filtering)) buffer_count计算如下： single_scan_read_cost计算如下： 1.0 + ((double)cache_record_length(join, idx) * prefix_rowcount / thd->variables.join_buff_size force_index模式: table->file->read_cost(tab->ref().key, 1,tab->records()) 见表五 非force_index模式: table->file->table_scan_cost() 见表五默认这个模式打开表五 引擎计算相关
场景值说明table->file->read_cost(index, ranges, rows)read_time(index, ranges, rows) *table->cost_model()->page_read_cost(1.0)index:index序号，ranges:range范围，rows:表行数table->file->table_scan_cost()(stats.data_file_length / IO_SIZE + 2) * table->cost_model()->page_read_cost(1.0)IO_SIZE=4096三、实际例子说明
接下来看一个例子来说明上面的代码。这里的例子不涉及keyuse_array数组有值的情况。
greatsql> SELECT * FROM t1 join t3 ON t1.c1=t3.ccc1 or t3.ccc1<3;
+----+------+---------------------+------+------+
| c1 | c2 | date1 | ccc1 | ccc2 |
+----+------+---------------------+------+------+
|1 | 10 | 2021-03-25 16:44:00 | 1 | aa1|
|5 | 5 | 2024-03-25 16:44:00 | 1 | aa1|
|3 | 4 | 2023-03-27 16:44:00 | 1 | aa1|
|2 | 1 | 2022-03-26 16:44:00 | 1 | aa1|
|1 | 10 | 2021-03-25 16:44:00 | 2 | bb1|
|5 | 5 | 2024-03-25 16:44:00 | 2 | bb1|
|3 | 4 | 2023-03-27 16:44:00 | 2 | bb1|
|2 | 1 | 2022-03-26 16:44:00 | 2 | bb1|
|3 | 4 | 2023-03-27 16:44:00 | 3 | cc1|
+----+------+---------------------+------+------+
greatsql> EXPLAIN FORMAT=TREE SELECT * FROM t1 join t3 ON t1.c1=t3.ccc1 or t3.ccc1<3;
+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| EXPLAIN |
+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| -> Filter: ((t3.ccc1 = t1.c1) or (t3.ccc1 < 3))(cost=3.65 rows=9)
-> Inner hash join (no condition)(cost=3.65 rows=9)
-> Table scan on t3(cost=0.12 rows=5)
-> Hash
-> Index scan on t1 using idx2(cost=1.40 rows=4)
|
+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
greatsql> SELECT * FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE;
"considered_execution_plans": [
{
"plan_prefix": [
],
"table": "`t1`",
"best_access_path": {
"considered_access_paths": [
{
"rows_to_scan": 4,
"filtering_effect": [
],
"final_filtering_effect": 1,
"access_type": "scan",
"resulting_rows": 4,# 这个值就是rows_after_filtering
"cost": 0.65, # 计算公式=read_cost(0.25)+cost_model->row_evaluate_cost(1 * rows_after_filtering)
"chosen": true
}
]
},
"condition_filtering_pct": 100, # 这个值就是filter_effect * 100
"rows_for_plan": 4, # 这个值就是prefix_rowcount=rows_to_scan * filter_effect / 100
"cost_for_plan": 0.65, # 这个值就是prefix_cost=read_cost(0.25) + cm->row_evaluate_cost(prefix_rowcount=4)
"rest_of_plan": [
{
"plan_prefix": [ # 当前左连接表为t1
"`t1`"
],
"table": "`t3`", # 右连接表为t3
"best_access_path": {
"considered_access_paths": [
{
"rows_to_scan": 5,
"filtering_effect": [
],
"final_filtering_effect": 1,
"access_type": "scan",
"using_join_cache": true,
"buffers_needed": 1,
"resulting_rows": 5, # 这个值就是rows_after_filtering
"cost": 2.25005, # 计算公式read_cost(0.25)+cost_model->row_evaluate_cost(prefix_rowcount * rows_after_filtering=4 * 5)
"chosen": true
}
]
},
"condition_filtering_pct": 46.664, # 这个值计算过程见下面<<附录：t3的filter_effect值计算>> ※bug？用explain看到的和直接敲命令看到的不一样，前者46.664后者为100，而"rows_for_plan"会变为没有过滤的20
"rows_for_plan": 9.3328, # 这个值就是prefix_rowcount=4(上一张表的prefix_rowcount) * 5 * 46.664 / 100
"cost_for_plan": 2.90005, # 这个值就是prefix_cost=0.65(上一张表的prefix_cost)+read_cost(0.25) + cm->row_evaluate_cost(20=4*5)
"chosen": true
}
]
},
{
"plan_prefix": [
],
"table": "`t3`",
"best_access_path": {
"considered_access_paths": [
{
"rows_to_scan": 5,
"filtering_effect": [
],
"final_filtering_effect": 1,
"access_type": "scan",
"resulting_rows": 5,
"cost": 0.75, # 计算公式read_cost(0.25)+5行*0.1
"chosen": true
}
]
},
"condition_filtering_pct": 100,
"rows_for_plan": 5, # 这个值就是prefix_rowcount
"cost_for_plan": 0.75, # 这个值就是prefix_cost
"rest_of_plan": [
{
"plan_prefix": [ # 当前左连接表为t3
"`t3`"
],
"table": "`t1`", # 右连接表为t1
"best_access_path": {
"considered_access_paths": [
{
"rows_to_scan": 4,
"filtering_effect": [
],
"final_filtering_effect": 1,
"access_type": "scan",
"using_join_cache": true,
"buffers_needed": 1,
"resulting_rows": 4,
"cost": 3.00776, # 计算公式read_cost(1.007)+cost_model->row_evaluate_cost(prefix_rowcount * rows_after_filtering=5 * 4)
"chosen": true
}
]
},
"condition_filtering_pct": 100,
"rows_for_plan": 20, # 这个值就是prefix_rowcount=5(上一张表的prefix_rowcount) * 4 * 100 / 100
"cost_for_plan": 3.75776, # 这个值就是prefix_cost=0.75(上一张表的prefix_cost)+read_cost(1.007) + cm->row_evaluate_cost(20=5*4)
"pruned_by_cost": true # 因为这里算出来的3.75776 > 2.90005，因此被裁剪掉了
}
]
}
]
},
{
"attaching_conditions_to_tables": { # 添加一些附加条件
"original_condition": "((`t3`.`ccc1` = `t1`.`c1`) or (`t3`.`ccc1` < 3))",
"attached_conditions_computation": [
],
"attached_conditions_summary": [ # t1作为驱动表，执行全表扫描，因此不需要任何条件
{
"table": "`t1`",
"attached": null
},
{
"table": "`t3`",
"attached": "((`t3`.`ccc1` = `t1`.`c1`) or (`t3`.`ccc1` < 3))" #t3作为连接表，按照条件过滤
}
]
}
},
{
"finalizing_table_conditions": [
{
"table": "`t3`",
"original_table_condition": "((`t3`.`ccc1` = `t1`.`c1`) or (`t3`.`ccc1` < 3))",
"final_table_condition ": "((`t3`.`ccc1` = `t1`.`c1`) or (`t3`.`ccc1` < 3))"
}
]
},
{
"refine_plan": [
{
"table": "`t1`"
},
{
"table": "`t3`"
}
]
}
]
}
},
{
"join_explain": {
"select#": 1,
"steps": [
]
}
}
]
} |
# 附录：t3的filter_effect值计算
# 这个函数计算t1.c1=t3.ccc1和t3.ccc1<3的filter_effect值，计算过程见下面
float Item_cond_or::get_filtering_effect(THD *thd, table_map filter_for_table,
table_map read_tables,
const MY_BITMAP *fields_to_ignore,
double rows_in_table) {
while ((item = it++)) {
const float cur_filter = item->get_filtering_effect(
thd, filter_for_table, read_tables, fields_to_ignore, rows_in_table);
# 第一次：计算t1.c1=t3.ccc1，返回 1/5=20%，其中5是总行数。
# 第二次：计算t3.ccc1<3，返回COND_FILTER_INEQUALITY=0.333
# 最后的filter=0.2+0.333 - (0.2*0.333)=0.4664
filter = filter + cur_filter - (filter * cur_filter);
}
return filter;
}现在把之前03的内容拿过来回顾一下当时的结果，看看为何是当时的结果。从下面结果可以看出，当存在keyuse的时候，优化器最后走哪条索引是通过find_best_ref()函数来决定的。
greatsql> EXPLAIN SELECT * FROM t1 join t2 ON t1.c1=t2.cc1 and t1.c1
页:
[1]