MySQL 用 limit 為什么會(huì)影響性能？
一，前言
首先說明一下MySQL的版本：
mysql> select version();
+-----------+
| version() |
+-----------+
| 5.7.17 |
+-----------+
1 row in set (0.00 sec)
表結(jié)構(gòu)：
mysql> desc test;
+--------+---------------------+------+-----+---------+----------------+
| Field | Type | Null | Key | Default | Extra |
+--------+---------------------+------+-----+---------+----------------+
| id | bigint(20) unsigned | NO | PRI | NULL | auto_increment |
| val | int(10) unsigned | NO | MUL | 0 | |
| source | int(10) unsigned | NO | | 0 | |
+--------+---------------------+------+-----+---------+----------------+
3 rows in set (0.00 sec)
id為自增主鍵，val為非唯一索引。
灌入大量數(shù)據(jù)，共500萬：
mysql> select count(*) from test;
+----------+
| count(*) |
+----------+
| 5242882 |
+----------+
1 row in set (4.25 sec)
我們知道，當(dāng)limit offset rows中的offset很大時(shí)，會(huì)出現(xiàn)效率問題：
mysql> select * from test where val=4 limit 300000,5;
+---------+-----+--------+
| id | val | source |
+---------+-----+--------+
| 3327622 | 4 | 4 |
| 3327632 | 4 | 4 |
| 3327642 | 4 | 4 |
| 3327652 | 4 | 4 |
| 3327662 | 4 | 4 |
+---------+-----+--------+
5 rows in set (15.98 sec)
為了達(dá)到相同的目的，我們一般會(huì)改寫成如下語(yǔ)句：
mysql> select * from test a inner join (select id from test where val=4 limit 300000,5) b on a.id=b.id;
+---------+-----+--------+---------+
| id | val | source | id |
+---------+-----+--------+---------+
| 3327622 | 4 | 4 | 3327622 |
| 3327632 | 4 | 4 | 3327632 |
| 3327642 | 4 | 4 | 3327642 |
| 3327652 | 4 | 4 | 3327652 |
| 3327662 | 4 | 4 | 3327662 |
+---------+-----+--------+---------+
5 rows in set (0.38 sec)
時(shí)間相差很明顯。
為什么會(huì)出現(xiàn)上面的結(jié)果？我們看一下select * from test where val=4 limit 300000,5;的查詢過程：
查詢到索引葉子節(jié)點(diǎn)數(shù)據(jù)。
根據(jù)葉子節(jié)點(diǎn)上的主鍵值去聚簇索引上查詢需要的全部字段值。
類似于下面這張圖：
像上面這樣，需要查詢300005次索引節(jié)點(diǎn)，查詢300005次聚簇索引的數(shù)據(jù)，最后再將結(jié)果過濾掉前300000條，取出最后5條。MySQL耗費(fèi)了大量隨機(jī)I/O在查詢聚簇索引的數(shù)據(jù)上，而有300000次隨機(jī)I/O查詢到的數(shù)據(jù)是不會(huì)出現(xiàn)在結(jié)果集當(dāng)中的。
肯定會(huì)有人問：既然一開始是利用索引的，為什么不先沿著索引葉子節(jié)點(diǎn)查詢到最后需要的5個(gè)節(jié)點(diǎn)，然后再去聚簇索引中查詢實(shí)際數(shù)據(jù)。這樣只需要5次隨機(jī)I/O，類似于下面圖片的過程：
其實(shí)我也想問這個(gè)問題。
證實(shí)
下面我們實(shí)際操作一下來證實(shí)上述的推論：
為了證實(shí)select * from test where val=4 limit 300000,5是掃描300005個(gè)索引節(jié)點(diǎn)和300005個(gè)聚簇索引上的數(shù)據(jù)節(jié)點(diǎn)，我們需要知道MySQL有沒有辦法統(tǒng)計(jì)在一個(gè)sql中通過索引節(jié)點(diǎn)查詢數(shù)據(jù)節(jié)點(diǎn)的次數(shù)。我先試了Handler_read_*系列，很遺憾沒有一個(gè)變量能滿足條件。
我只能通過間接的方式來證實(shí)：
InnoDB中有buffer pool。里面存有最近訪問過的數(shù)據(jù)頁(yè)，包括數(shù)據(jù)頁(yè)和索引頁(yè)。所以我們需要運(yùn)行兩個(gè)sql，來比較buffer pool中的數(shù)據(jù)頁(yè)的數(shù)量。預(yù)測(cè)結(jié)果是運(yùn)行select * from test a inner join (select id from test where val=4 limit 300000,5) b>之后，buffer pool中的數(shù)據(jù)頁(yè)的數(shù)量遠(yuǎn)遠(yuǎn)少于select * from test where val=4 limit 300000,5;
select * from test where val=4 limit 300000,5
mysql> select index_name,count(*) from information_schema.INNODB_BUFFER_PAGE where INDEX_NAME in('val','primary') and TABLE_NAME like '%test%' group by index_name;
Empty set (0.04 sec)可以看出，目前buffer pool中沒有關(guān)于test表的數(shù)據(jù)頁(yè)。
mysql> select * from test where val=4 limit 300000,5;
+---------+-----+--------+
| id | val | source |
+---------+-----+--------+
| 3327622 | 4 | 4 |
| 3327632 | 4 | 4 |
| 3327642 | 4 | 4 |
| 3327652 | 4 | 4 |
| 3327662 | 4 | 4 |
+---------+-----+--------+
5 rows in set (26.19 sec)
mysql> select index_name,count(*) from information_schema.INNODB_BUFFER_PAGE where INDEX_NAME in('val','primary') and TABLE_NAME like '%test%' group by index_name;
+------------+----------+
| index_name | count(*) |
+------------+----------+
| PRIMARY | 4098 |
| val | 208 |
+------------+----------+
2 rows in set (0.04 sec)可以看出，此時(shí)buffer pool中關(guān)于test表有4098個(gè)數(shù)據(jù)頁(yè)，208個(gè)索引頁(yè)。
select * from test a inner join (select id from test where val=4 limit 300000,5) b>為了防止上次試驗(yàn)的影響，我們需要清空buffer pool，重啟mysql。mysqladmin shutdown
/usr/local/bin/mysqld_safe &
mysql> select index_name,count(*) from information_schema.INNODB_BUFFER_PAGE where INDEX_NAME in('val','primary') and TABLE_NAME like '%test%' group by index_name;
Empty set (0.03 sec)
運(yùn)行sql： mysql> select * from test a inner join (select id from test where val=4 limit 300000,5) b on a.id=b.id;
+---------+-----+--------+---------+
| id | val | source | id |
+---------+-----+--------+---------+
| 3327622 | 4 | 4 | 3327622 |
| 3327632 | 4 | 4 | 3327632 |
| 3327642 | 4 | 4 | 3327642 |
| 3327652 | 4 | 4 | 3327652 |
| 3327662 | 4 | 4 | 3327662 |
+---------+-----+--------+---------+
5 rows in set (0.09 sec)
mysql> select index_name,count(*) from information_schema.INNODB_BUFFER_PAGE where INDEX_NAME in('val','primary') and TABLE_NAME like '%test%' group by index_name;
+------------+----------+
| index_name | count(*) |
+------------+----------+
| PRIMARY | 5 |
| val | 390 |
+------------+----------+
2 rows in set (0.03 sec)
我們可以看明顯的看出兩者的差別：第一個(gè)sql加載了4098個(gè)數(shù)據(jù)頁(yè)到buffer pool，而第二個(gè)sql只加載了5個(gè)數(shù)據(jù)頁(yè)到buffer pool。符合我們的預(yù)測(cè)。也證實(shí)了為什么第一個(gè)sql會(huì)慢：讀取大量的無用數(shù)據(jù)行（300000），最后卻拋棄掉。
而且這會(huì)造成一個(gè)問題：加載了很多熱點(diǎn)不是很高的數(shù)據(jù)頁(yè)到buffer pool，會(huì)造成buffer pool的污染，占用buffer pool的空間。
遇到的問題
mysqladmin shutdown
/usr/local/bin/mysqld_safe &
mysql> select index_name,count(*) from information_schema.INNODB_BUFFER_PAGE where INDEX_NAME in('val','primary') and TABLE_NAME like '%test%' group by index_name;
Empty set (0.03 sec)
mysql> select * from test a inner join (select id from test where val=4 limit 300000,5) b on a.id=b.id;
+---------+-----+--------+---------+
| id | val | source | id |
+---------+-----+--------+---------+
| 3327622 | 4 | 4 | 3327622 |
| 3327632 | 4 | 4 | 3327632 |
| 3327642 | 4 | 4 | 3327642 |
| 3327652 | 4 | 4 | 3327652 |
| 3327662 | 4 | 4 | 3327662 |
+---------+-----+--------+---------+
5 rows in set (0.09 sec)
mysql> select index_name,count(*) from information_schema.INNODB_BUFFER_PAGE where INDEX_NAME in('val','primary') and TABLE_NAME like '%test%' group by index_name;
+------------+----------+
| index_name | count(*) |
+------------+----------+
| PRIMARY | 5 |
| val | 390 |
+------------+----------+
2 rows in set (0.03 sec)
而且這會(huì)造成一個(gè)問題：加載了很多熱點(diǎn)不是很高的數(shù)據(jù)頁(yè)到buffer pool，會(huì)造成buffer pool的污染，占用buffer pool的空間。
遇到的問題
為了在每次重啟時(shí)確保清空buffer pool，我們需要關(guān)閉innodb_buffer_pool_dump_at_shutdown和innodb_buffer_pool_load_at_startup，這兩個(gè)選項(xiàng)能夠控制數(shù)據(jù)庫(kù)關(guān)閉時(shí)dump出buffer pool中的數(shù)據(jù)和在數(shù)據(jù)庫(kù)開啟時(shí)載入在磁盤上備份buffer pool的數(shù)據(jù)。
參考資料：
1.https://explainextended.com/2009/10/23/mysql-order-by-limit-performance-late-row-lookups/
2.https://dev.mysql.com/doc/refman/5.7/en/innodb-information-schema-buffer-pool-tables.html