在 MySql 建表時(shí)候一般會(huì)指定字符集，大多數(shù)情況下為了更好的兼容性無(wú)腦選了 utf8mb4。但是有時(shí)會(huì)因?yàn)檫x錯(cuò)，或歷史遺留問(wèn)題，導(dǎo)致使用了 utf8 字符集。當(dāng)兩個(gè)表的字符集不一樣，在使用字符型字段進(jìn)行表連接查詢時(shí)，就需要特別注意下查詢耗時(shí)是否符合預(yù)期。
有次使用 left join 寫(xiě)一個(gè) SQL，發(fā)現(xiàn)用時(shí)明顯超過(guò)預(yù)期，經(jīng)過(guò)一頓折騰才發(fā)現(xiàn)是兩個(gè)表字符集不一樣，特此記錄一下。
問(wèn)題分析
mysql> SELECT COUNT( *) from app_bind_rel t left join app_config_control_sn p on t.host_sn = p.host_sn ;
+-----------+
| COUNT( *) |
+-----------+
| 13447 |
+-----------+
1 row in set (0.89 sec)
例如上面的 SQL，左表 1W 條數(shù)據(jù)，右表 400 多條數(shù)據(jù)，在 host_sn 字段上都有索引，查詢竟然用了近 900ms，怎么會(huì)這么慢？
mysql> explain SELECT COUNT( *) from app_bind_rel t left join app_config_control_sn p on t.host_sn = p.host_sn ;
+----+-------------+-------+------------+-------+---------------+-------------+---------+------+-------+----------+-----------------------------------------------------------------+
| id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra |
+----+-------------+-------+------------+-------+---------------+-------------+---------+------+-------+----------+-----------------------------------------------------------------+
| 1 | SIMPLE | t | NULL | index | NULL | idx_host_sn | 122 | NULL | 10791 | 100.00 | Using index |
| 1 | SIMPLE | p | NULL | index | NULL | idx_host_sn | 152 | NULL | 457 | 100.00 | Using where; Using index; Using join buffer (Block Nested Loop) |
+----+-------------+-------+------------+-------+---------------+-------------+---------+------+-------+----------+-----------------------------------------------------------------+
2 rows in set, 1 warning (0.00 sec)
查看下執(zhí)行計(jì)劃，的確是使用了索引，但是細(xì)看 Extra 列發(fā)現(xiàn)較正常的連表查詢多了“Using join buffer (Block Nested Loop)”這一信息，這個(gè)具體是什么意思我們后面再說(shuō)。
然后我們?cè)倏聪略敿?xì)的執(zhí)行計(jì)劃，使用 explain formart=json。
{
"query_block": {
"select_id": 1,
"cost_info": {
"query_cost": "988640.52"
},
"nested_loop": [
{
"table": {
"table_name": "t",
"access_type": "index",
"key": "idx_host_sn",
"used_key_parts": [
"host_sn"
],
"key_length": "122",
"rows_examined_per_scan": 10791,
"rows_produced_per_join": 10791,
"filtered": "100.00",
"using_index": true,
"cost_info": {
"read_cost": "161.00",
"eval_cost": "2158.20",
"prefix_cost": "2319.20",
"data_read_per_join": "2M"
},
"used_columns": [
"host_sn"
]
}
},
{
"table": {
"table_name": "p",
"access_type": "index",
"key": "idx_host_sn",
"used_key_parts": [
"host_sn"
],
"key_length": "152",
"rows_examined_per_scan": 457,
"rows_produced_per_join": 4931487,
"filtered": "100.00",
"using_index": true,
"using_join_buffer": "Block Nested Loop",
"cost_info": {
"read_cost": "23.92",
"eval_cost": "986297.40",
"prefix_cost": "988640.52",
"data_read_per_join": "865M"
},
"used_columns": [
"host_sn"
],
"attached_condition": "<if>(is_not_null_compl(p), (`db0`.`t`.`host_sn` = convert(`db0`.`p`.`host_sn` using utf8mb4)), true)"
}
}
]
}
}
特別需要關(guān)注的是這一對(duì) KV
"attached_condition": "<if>(is_not_null_compl(p), (`collection_bullet_0000`.`t`.`host_sn` = convert(`collection_bullet_0000`.`p`.`host_sn` using utf8mb4)), true)"
看字面意思就是當(dāng) p 表不為空的時(shí)候，執(zhí)行表連接需要先將 p 表的 host_sn 字段轉(zhuǎn)變?yōu)?utf8mb4 字符集。我們應(yīng)該都知道在表連接中使用了函數(shù)的話，是無(wú)法使用索引的。
所以再回到上面我看到的“Using join buffer (Block Nested Loop)”問(wèn)題，來(lái)解釋下這是一個(gè)什么過(guò)程。
Nested-Loop Join
MySql 官網(wǎng)對(duì) Nested-Loop Join 有做過(guò)解釋?zhuān)鋵?shí)做開(kāi)發(fā)的同學(xué)看到名字就大體知道是啥，不就是循環(huán)嵌套嘛。
MySql 中分為 Nested-Loop Join 算法跟 Block Nested-Loop Join 算法。
例如，有如下三個(gè)表，t1、t2、t3 使用了這三種 join type。
Table Join Type
t1 range
t2 ref
t3 ALL
當(dāng)使用 Nested-Loop Join 算法時(shí)，其 join 過(guò)程如下所示，其實(shí)就是簡(jiǎn)單的三層循環(huán)。
for each row in t1 matching range {
for each row in t2 matching reference key {
for each row in t3 {
if row satisfies join conditions, send to client
}
}
}
Block Nested-Loop Join（BNL） 算法是對(duì) Nested-Loop Join 算法的一種優(yōu)化。BNL 算法緩沖外部循環(huán)中讀取的行來(lái)減少內(nèi)部循環(huán)中讀取表的次數(shù)。例如，將 10 行數(shù)據(jù)讀取到緩沖器中，并且將緩沖器傳遞到下一個(gè)循環(huán)內(nèi)部，內(nèi)部循環(huán)中讀取的每一行與緩沖器中的所有 10 行進(jìn)行比較。這將使讀取內(nèi)部表的次數(shù)減少一個(gè)數(shù)量級(jí)。
for each row in t1 matching range {
for each row in t2 matching reference key {
store used columns from t1, t2 in join buffer
if buffer is full {
for each row in t3 {
for each t1, t2 combination in join buffer {
if row satisfies join conditions, send to client
}
}
empty join buffer
}
}
}
if buffer is not empty {
for each row in t3 {
for each t1, t2 combination in join buffer {
if row satisfies join conditions, send to client
}
}
}
算法實(shí)現(xiàn)如上，只有當(dāng) “join buffer” 滿的時(shí)候才會(huì)觸發(fā) t3 表的讀取，如果 “join buffer” 的 size = 10 那么就可以減少 10 倍的 t3 表被讀取次數(shù)，從內(nèi)存中讀取數(shù)據(jù)的效率顯然要比從磁盤(pán)讀取的效率高的多。從而提升 join 的效率。
但其實(shí)再好的優(yōu)化畢竟也是嵌套循環(huán)，做開(kāi)發(fā)的同學(xué)應(yīng)該都知道 O(N2) 的時(shí)間復(fù)雜度是無(wú)法接受的。這也是我們這個(gè)查詢這么慢的根因。
解決辦法
解決辦法其實(shí)很簡(jiǎn)單，修改右表的字符集就可以解決。
在變更數(shù)據(jù)集之前我們先用 show table status 查看下當(dāng)前表的狀態(tài)。
mysql> show table status like 'app_config_control_sn';
+-----------------------+--------+---------+------------+------+----------------+-------------+-----------------+--------------+-----------+----------------+---------------------+---------------------+------------+-----------------+----------+----------------+---------+
| Name | Engine | Version | Row_format | Rows | Avg_row_length | Data_length | Max_data_length | Index_length | Data_free | Auto_increment | Create_time | Update_time | Check_time | Collation | Checksum | Create_options | Comment |
+-----------------------+--------+---------+------------+------+----------------+-------------+-----------------+--------------+-----------+----------------+---------------------+---------------------+------------+-----------------+----------+----------------+---------+
| app_config_control_sn | InnoDB | 10 | Dynamic | 457 | 143 | 65536 | 0 | 32768 | 0 | 1041 | 2023-04-17 03:25:45 | 2023-04-17 03:27:24 | NULL | utf8_general_ci | NULL | | SN |
+-----------------------+--------+---------+------------+------+----------------+-------------+-----------------+--------------+-----------+----------------+---------------------+---------------------+------------+-----------------+----------+----------------+---------+
1 row in set (0.00 sec)
接著使用如下命令變更表的字符集。
mysql> ALTER TABLE app_config_control_sn CONVERT TO CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci;
Query OK, 457 rows affected (0.09 sec)
Records: 457 Duplicates: 0 Warnings: 0
再次使用 show table status 命令查看下表的狀態(tài)。
mysql> show table status like 'app_config_control_sn';
+-----------------------+--------+---------+------------+------+----------------+-------------+-----------------+--------------+-----------+----------------+---------------------+---------------------+------------+--------------------+----------+----------------+---------+
| Name | Engine | Version | Row_format | Rows | Avg_row_length | Data_length | Max_data_length | Index_length | Data_free | Auto_increment | Create_time | Update_time | Check_time | Collation | Checksum | Create_options | Comment |
+-----------------------+--------+---------+------------+------+----------------+-------------+-----------------+--------------+-----------+----------------+---------------------+---------------------+------------+--------------------+----------+----------------+---------+
| app_config_control_sn | InnoDB | 10 | Dynamic | 457 | 143 | 65536 | 0 | 32768 | 0 | 1041 | 2023-04-17 03:50:11 | 2023-04-17 03:50:11 | NULL | utf8mb4_general_ci | NULL | | SN |
+-----------------------+--------+---------+------------+------+----------------+-------------+-----------------+--------------+-----------+----------------+---------------------+---------------------+------------+--------------------+----------+----------------+---------+
1 row in set (0.01 sec)
可以看到表的字符集已經(jīng)發(fā)生了變化，那我們?cè)俅螆?zhí)行開(kāi)始的 SQL 及 explain 語(yǔ)句，確認(rèn)下問(wèn)題是否已經(jīng)解決。
mysql> SELECT COUNT( *) from app_bind_rel t left join app_config_control_sn p on t.host_sn = p.host_sn ;
+-----------+
| COUNT( *) |
+-----------+
| 13447 |
+-----------+
1 row in set (0.03 sec)
mysql> explain SELECT COUNT( *) from app_bind_rel t left join app_config_control_sn p on t.host_sn = p.host_sn ;
+----+-------------+-------+------------+-------+---------------+-------------+---------+---------------+-------+----------+--------------------------+
| id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra |
+----+-------------+-------+------------+-------+---------------+-------------+---------+---------------+-------+----------+--------------------------+
| 1 | SIMPLE | t | NULL | index | NULL | idx_host_sn | 122 | NULL | 10791 | 100.00 | Using index |
| 1 | SIMPLE | p | NULL | ref | idx_host_sn | idx_host_sn | 202 | db0.t.host_sn | 2 | 100.00 | Using where; Using index |
+----+-------------+-------+------------+-------+---------------+-------------+---------+---------------+-------+----------+--------------------------+
2 rows in set, 1 warning (0.00 sec)
可以看到耗時(shí)已經(jīng)只需要 30ms 左右，這個(gè)就比較符合預(yù)期，而在執(zhí)行計(jì)劃中也不再會(huì)有“Using join buffer (Block Nested Loop)”信息。
其他
mysql> SELECT COUNT( *) from app_bind_rel t join app_config_control_sn p on t.host_sn = p.host_sn ;
+-----------+
| COUNT( *) |
+-----------+
| 730 |
+-----------+
1 row in set (0.01 sec)
在沒(méi)有變更字符集之前，當(dāng)我們將 left join 修改為 join 的時(shí)候會(huì)發(fā)現(xiàn)耗時(shí)減少了 100 倍，只用了 10 ms，這是為什么呢？
mysql> explain SELECT COUNT( *) from app_bind_rel t join app_config_control_sn p on t.host_sn = p.host_sn ;
+----+-------------+-------+------------+-------+---------------+-------------+---------+------+------+----------+--------------------------+
| id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra |
+----+-------------+-------+------------+-------+---------------+-------------+---------+------+------+----------+--------------------------+
| 1 | SIMPLE | p | NULL | index | NULL | idx_host_sn | 152 | NULL | 457 | 100.00 | Using index |
| 1 | SIMPLE | t | NULL | ref | idx_host_sn | idx_host_sn | 122 | func | 1 | 100.00 | Using where; Using index |
+----+-------------+-------+------------+-------+---------------+-------------+---------+------+------+----------+--------------------------+
2 rows in set, 1 warning (0.00 sec)
查看執(zhí)行計(jì)劃，發(fā)現(xiàn)使用 join 的時(shí)候不會(huì)有 “Using join buffer (Block Nested Loop)”。再細(xì)看執(zhí)行計(jì)劃，發(fā)現(xiàn)驅(qū)動(dòng)表已經(jīng)由 t 表變?yōu)榱?p 表。
{
"query_block": {
"select_id": 1,
"cost_info": {
"query_cost": "643.80"
},
"nested_loop": [
{
"table": {
"table_name": "p",
"access_type": "index",
"key": "idx_host_sn",
"used_key_parts": [
"host_sn"
],
"key_length": "152",
"rows_examined_per_scan": 457,
"rows_produced_per_join": 457,
"filtered": "100.00",
"using_index": true,
"cost_info": {
"read_cost": "4.00",
"eval_cost": "91.40",
"prefix_cost": "95.40",
"data_read_per_join": "82K"
},
"used_columns": [
"host_sn"
]
}
},
{
"table": {
"table_name": "t",
"access_type": "ref",
"possible_keys": [
"idx_host_sn"
],
"key": "idx_host_sn",
"used_key_parts": [
"host_sn"
],
"key_length": "122",
"ref": [
"func"
],
"rows_examined_per_scan": 1,
"rows_produced_per_join": 457,
"filtered": "100.00",
"using_index": true,
"cost_info": {
"read_cost": "457.00",
"eval_cost": "91.40",
"prefix_cost": "643.80",
"data_read_per_join": "117K"
},
"used_columns": [
"host_sn"
],
"attached_condition": "(`db0`.`t`.`host_sn` = convert(`db0`.`p`.`host_sn` using utf8mb4))"
}
}
]
}
}
查看詳細(xì)的執(zhí)行計(jì)劃，可以看到
"attached_condition": "(`collection_bullet_0000`.`t`.`host_sn` = convert(`collection_bullet_0000`.`p`.`host_sn` using utf8mb4))"
這對(duì) KV 依然是存在的，但是 "using_join_buffer": "Block Nested Loop" 已經(jīng)不存在了。這個(gè)其實(shí)主要是因?yàn)楫?dāng) p 表變?yōu)轵?qū)動(dòng)表的時(shí)候，會(huì)先將自己的 host_sn 字段轉(zhuǎn)為 utf8mb4 字符集，再與 t 表進(jìn)行關(guān)聯(lián)。t 表由于本來(lái)就是 utf8mb4 字符集且存在索引，就可以正常走數(shù)據(jù)庫(kù)索引了，所以查詢耗時(shí)也就大大降低。而使用 left join 時(shí)候，t 表作為驅(qū)動(dòng)表是無(wú)法優(yōu)化改變的。
可見(jiàn)在表連接中即使使用了函數(shù)也不一定就沒(méi)法走索引，關(guān)鍵還是要看用法及明確處理過(guò)程。
記得剛學(xué)習(xí)數(shù)據(jù)庫(kù)的時(shí)候，老師還特別強(qiáng)調(diào)驅(qū)動(dòng)表一定要寫(xiě)在左邊，而隨著數(shù)據(jù)庫(kù)技術(shù)的不斷迭代發(fā)展，數(shù)據(jù)庫(kù)已經(jīng)能更智能的自動(dòng)幫我們優(yōu)化處理過(guò)程，之前很多的數(shù)據(jù)庫(kù)規(guī)則也不需要了。
