MySQL实战系列:大字段如何优化

  1. 云栖社区>
  2. 博客>
  3. 正文

MySQL实战系列:大字段如何优化

兰春 2016-08-19 09:32:49 浏览4563
展开阅读全文

背景

线上发现一张表,1亿的数据量,物理大小尽然惊人的大,1.2T
最终发现,原来有很多字段,10个varchar,1个text
这么大的表,会给运维带来很大的痛苦:DDL咋办?恢复咋办?备份咋办?

基本知识:InnoDB Storage Architecture for InnoDB On Disk Format

蓝图: database --> tablespaces --> pages --> rows --> columns

InnoDB 物理结构存储结构

database_file

InnoDB 逻辑存储结构

tablespace

InnoDB page 存储结构

页类型

数据页(B-tree Node)
undo页(undo Log Page)
系统页(System Page)
事务数据页(Transaction system Page)
插入缓冲位图页(Insert Buffer Page)
未压缩的二进制大对象页(Uncompressd BLOB Page)
压缩的二进制大对象页(compressd BLOB Page)

页大小

默认16k(若果没有特殊情况,下面介绍的都是默认16k大小为准)
一个页内必须存储2行记录,否则就不是B+tree,而是链表了

结构图

innodb_page

InnoDB row 存储结构

rows 文件格式总体规划图

row_file_format

row-fomat为Compact的结构图

compact

row-fomat为Redundant的结构图

不常用

compress & dynamic 与 Compact 的区别之处

dynamic

字段之字符串类型

char(N) vs varchar(N)

不管是char,还是varchar,在compact row-format格式下,NULL都不占用任何存储空间
在多字节字符集的情况下,CHAR vs VARCHAR 的实际行存储基本没区别
CHAR不管是否是多字符集,对未能占满长度的字符还是会填充0x20
规范中:对char和varchar可以不做要求

varchar(N) : 255 vs 256

当实际长度大于255的时候,变长字段长度列表需要用两个字节存储,也就意味着每一行数据都会增加1个字节
实测下来存储空间增长并不算大,且性能影响也不大,所以,尽量在256之内吧

varchar(N) & char(N) 的最大限制

char的最大限制是: N<=255
varchar 的最大限制是: N<=65535 , 注意官方文档说的是N是字节,并且说的是一行的所有字段的总和小于65535,而varchar(N)中的N表示的是字符。
测试后发现,65535并不是最大限制,最大的限制是65532

[MySQL 5.6.27]

* char的最大限制是: N<=255
root:test> create table test( a char(65535))charset=latin1 engine=innodb;
ERROR 1074 (42000): Column length too big for column 'a' (max = 255); use BLOB or TEXT instead

* 测试后发现,65535并不是最大限制,最大的限制是65532
root:test> create table test( a varchar(65535))charset=latin1 engine=innodb;
ERROR 1118 (42000): Row size too large. The maximum row size for the used table type, not counting BLOBs, is 65535. This includes storage overhead, check the manual. You have to change some columns to TEXT or BLOBs

root:test> create table test( a varchar(65532))charset=latin1 engine=innodb;
Query OK, 0 rows affected (0.00 sec)

* varchar 的最大限制是: N<=65535 , 注意官方文档说的是N是字节,并且说的是一行的所有字段的总和小于65535,而varchar(N)中的N表示的是字符

root:test> create table test_1( a varchar(30000),b varchar(30000),c varchar(5535))charset=latin1 engine=innodb;
ERROR 1118 (42000): Row size too large. The maximum row size for the used table type, not counting BLOBs, is 65535. This includes storage overhead, check the manual. You have to change some columns to TEXT or BLOBs

* varchar(N)中的N表示的是字符

root:test> create table test_1( a varchar(50000))charset=utf8 engine=innodb;
Query OK, 0 rows affected, 1 warning (0.00 sec)

root:test> show warnings;
+-------+------+--------------------------------------------+
| Level | Code | Message                                    |
+-------+------+--------------------------------------------+
| Note  | 1246 | Converting column 'a' from VARCHAR to TEXT |
+-------+------+--------------------------------------------+
1 row in set (0.00 sec)

root:test> show create table test_1;
+--------+-------------------------------------------------------------------------------+
| Table  | Create Table                                                                  |
+--------+-------------------------------------------------------------------------------+
| test_1 | CREATE TABLE `test_1` (
  `a` mediumtext
) ENGINE=InnoDB DEFAULT CHARSET=utf8 |
+--------+-------------------------------------------------------------------------------+
1 row in set (0.00 sec)

off-page: 行溢出

  • 为什么会有行溢出off-page这个概念呢

假设创建了一张表,里面有一个字段是a varchar(30000) , innoDB的页才16384个字节,如何存储的下呢?所以行溢出就来了嘛

  • 如何看出行溢出了?

可以通过姜承尧写的工具查看
其中溢出的页有 Uncompressed BLOB Page: 243453

[root()@xx script]# python py_innodb_page_info.py t.ibd

Total number of page: 537344:
Insert Buffer Bitmap: 33
Freshly Allocated Page: 74040
File Segment inode: 1
B-tree Node: 219784
File Space Header: 1
扩展描述页: 32
Uncompressed BLOB Page: 243453
  • 溢出有什么危害

溢出的数据不再存储在B+tree中
溢出的数据使用的是uncompress BLOB page,并且存储独享,这就是存储越来越大的真正原因
通过下面的测试,你会发现,t_long 插入的数据仅仅比 t_short 多了几个字节,但是最终的存储却是2~3倍的差距

* 表结构

root:test> show create table t_long;
+--------+---------------------------------------------------------------------------------------------------------+
| Table  | Create Table                                                                                            |
+--------+---------------------------------------------------------------------------------------------------------+
| t_long | CREATE TABLE `t_long` (
  `id` int(11) DEFAULT NULL,
  `col1` text
) ENGINE=InnoDB DEFAULT CHARSET=utf8 |
+--------+---------------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)

root:test> show create table t_short;
+---------+----------------------------------------------------------------------------------------------------------+
| Table   | Create Table                                                                                             |
+---------+----------------------------------------------------------------------------------------------------------+
| t_short | CREATE TABLE `t_short` (
  `id` int(11) DEFAULT NULL,
  `col1` text
) ENGINE=InnoDB DEFAULT CHARSET=utf8 |
+---------+----------------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)


* 测试案例

foreach $num (1 .. 48849){

        $sql_1 = "insert into $table_short select $num,repeat('a',8090)";
        $sql_2 = "insert into $table_long select $num,repeat('a',8098)";
        `$cmd -e " $sql_1 "`;
        `$cmd -e " $sql_2 "`;
}


* 最终的记录数

root:test> select count(*) from t_short;
+----------+
| count(*) |
+----------+
|    48849 |
+----------+
1 row in set (0.03 sec)

root:test> select count(*) from t_long;
+----------+
| count(*) |
+----------+
|    48849 |
+----------+
1 row in set (0.02 sec)


* 页类型的比较

[root()@xx script]# python py_innodb_page_info.py /data/mysql_data/test/t_short.ibd
Total number of page: 25344:
Insert Buffer Bitmap: 2
Freshly Allocated Page: 887
File Segment inode: 1
B-tree Node: 24452
File Space Header: 1
扩展描述页: 1



[root()@xx script]# python py_innodb_page_info.py /data/mysql_data/test/t_long.ibd
Total number of page: 60160:
Insert Buffer Bitmap: 4
Freshly Allocated Page: 8582
File Segment inode: 1
B-tree Node: 2720
File Space Header: 1
扩展描述页: 3
Uncompressed BLOB Page: 48849


* 最终大小的对比

[root()@xx test]# du -sh * | grep 'long\|short' | grep ibd
941M    t_long.ibd
397M    t_short.ibd

* 结论

t_short 的表,在400M左右可以理解,因为 8k * 48849 = 400M

t_long 的表,由于独享48849个Uncompressed BLOB Page,严重浪费空间
  • 什么情况下会溢出

原则:只要一行记录的总和超过8k,就会溢出。
所以:varchar(9000) 或者 varchar(3000) + varchar(3000) + varchar(3000),当实际长度大于8k的时候,就会溢出
所以:Blob,text,一行数据如果实际长度大于8k会溢出,如果实际长度小于8k则不会溢出,并非所有的blob,text都会溢出

  • 多列总和大字段 vs 一列大字段

多个大字段会导致多次off-page

root:test> show create table t_3_col;
+---------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------+
| Table   | Create Table
                                       |
+---------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------+
| t_3_col | CREATE TABLE `t_3_col` (
  `id` int(11) DEFAULT NULL,
  `col1` varchar(7000) DEFAULT NULL,
  `col2` varchar(7000) DEFAULT NULL,
  `col3` varchar(7000) DEFAULT NULL
) ENGINE=InnoDB DEFAULT CHARSET=utf8 |
+---------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------+
1 row in set (0.00 sec)

root:test> show create table t_1_col;
+---------+---------------------------------------------------------------------------------------------------------------------------------+
| Table   | Create Table                                                                                                                    |
+---------+---------------------------------------------------------------------------------------------------------------------------------+
| t_1_col | CREATE TABLE `t_1_col` (
  `id` int(11) DEFAULT NULL,
  `col1` varchar(21000) DEFAULT NULL
) ENGINE=InnoDB DEFAULT CHARSET=utf8 |
+---------+---------------------------------------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)

root:test>
root:test>
root:test> insert into t_1_col(col1) select repeat('a',21000);
Query OK, 1 row affected (0.00 sec)
Records: 1  Duplicates: 0  Warnings: 0

root:test>
root:test>
root:test> insert into t_3_col(col1,col2,col3) select repeat('a',7000),repeat('a',7000),repeat('a',7000);
Query OK, 1 row affected (0.00 sec)
Records: 1  Duplicates: 0  Warnings: 0


[root()@xx script]# python py_innodb_page_info.py /data/mysql_data/test/t_1_col.ibd
Total number of page: 6:
Insert Buffer Bitmap: 1
Uncompressed BLOB Page: 2
File Space Header: 1
B-tree Node: 1
File Segment inode: 1

[root()@xx script]# python py_innodb_page_info.py /data/mysql_data/test/t_3_col.ibd
Total number of page: 7:
Insert Buffer Bitmap: 1
Uncompressed BLOB Page: 3
File Space Header: 1
B-tree Node: 1
File Segment inode: 1

如何对大字段进行优化

如果有多个大字段,尽量序列化后,存储在同一列中,避免多次off-page
将text等大字段从主表中拆分出来,a)存储到key-value中 b)存储在单独的一张子表中,并且压缩
必须保证一行记录小于8k

参考

http://dev.mysql.com/doc/refman/5.7/en/innodb-storage-engine.html
INNOBASE 官方文档
MySQL技术内幕 InnoDB存储引擎 --姜承尧
http://hedengcheng.com/ --何登成
http://imysql.cn/ --叶金荣

网友评论

登录后评论
0/500
评论
兰春
+ 关注