开发指南—DDL语句—分库分表语法—CREATE TABLE

本文涉及的产品
云原生数据库 PolarDB 分布式版,标准版 2核8GB
简介: 本文主要介绍使用DDL语句进行建表的语法、子句、参数和基本方式。

语法


CREATE [SHADOW] TABLE [IF NOT EXISTS] tbl_name
    (create_definition, ...)
    [table_options]
    [drds_partition_options]
create_definition:
    col_name column_definition
  | mysql_create_definition
  | [UNIQUE] GLOBAL INDEX index_name [index_type] (index_sharding_col_name,...)
      [global_secondary_index_option]
      [index_option] ...
# 全局二级索引相关
global_secondary_index_option:
    [COVERING (col_name,...)]
    [drds_partition_options]
# 分库分表子句
drds_partition_options:
    DBPARTITION BY db_partition_algorithm
    [TBPARTITION BY table_partition_algorithm [TBPARTITIONS num]]
    [LOCALITY=locality_option]
db_sharding_algorithm:
    HASH([col_name])
  | {YYYYMM|YYYYWEEK|YYYYDD|YYYYMM_OPT|YYYYWEEK_OPT|YYYYDD_OPT}(col_name)
  | UNI_HASH(col_name)
  | RIGHT_SHIFT(col_name, n)
  | RANGE_HASH(col_name, col_name, n)
table_sharding_algorithm:
    HASH(col_name)
  | {MM|DD|WEEK|MMDD|YYYYMM|YYYYWEEK|YYYYDD|YYYYMM_OPT|YYYYWEEK_OPT|YYYYDD_OPT}(col_name)
  | UNI_HASH(col_name)
  | RIGHT_SHIFT(col_name, n)
  | RANGE_HASH(col_name, col_name, n)
# 可以在创建单表时指定该表的存储位置
locality_option:
    'dn=storage_inst_id_list'
storage_inst_id_list:
    storage_inst_id[,storage_inst_id_list]
# 以下为MySQL DDL语法
index_sharding_col_name:
    col_name [(length)] [ASC | DESC]
index_option:
    KEY_BLOCK_SIZE [=] value
  | index_type
  | WITH PARSER parser_name
  | COMMENT 'string'
index_type:
    USING {BTREE | HASH}


说明 PolarDB-X DDL语法基于MySQL语法,以上主要列出了差异部分,详细语法请参见MySQL 文档

分库分表子句和参数

  • DBPARTITION BY hash(partition_key):指定分库键和分库算法。
  • TBPARTITION BY { HASH(column) | {MM|DD|WEEK|MMDD|YYYYMM|YYYYWEEK|YYYYDD|YYYYMM_OPT|YYYYWEEK_OPT|YYYYDD_OPT}(column)(可选):默认与DBPARTITION BY相同,指定物理表使用什么方式映射数据。
  • TBPARTITIONS num(可选):每个库上的物理表数目(默认为1),如不分表,就不需要指定该字段。
  • 拆分函数的详细介绍,请参见拆分函数

全局二级索引定义子句

  • [UNIQUE] GLOBAL:定义全局二级索引,UNIQUE GLOBAL代表全局唯一索引。
  • index_name:索引名,也是索引表的名称。
  • index_type:索引表中分库分表键上局部索引的类型,支持范围请参见MySQL 文档
  • index_sharding_col_name,...:索引列,包含且仅包含索引表的全部分库分表键。详情请参见全局二级索引
  • global_secondary_index_option:全局二级索引的扩展语法。
  • COVERING (col_name,...):覆盖列,索引表中除索引列以外的其他列,默认包含主键和主表的分库分表键。详情请参见全局二级索引
  • drds_partition_options:索引表的分库分表子句,详情请参见分库分表子句和参数
  • index_option:索引表中分库分表键上局部索引的属性,详情请参见MySQL 文档

全链路压测影子表子句

SHADOW:创建全链路压测影子表,表名必须以_test_为前缀,前缀后的表名部分必须与关联的正式表名一致,且正式表必须先于影子表创建。

LOCALITY

通过LOCALITY指定单表的存储位置。


说明

  • 通过LOCALITY语法指定了数据库或单表的位置之后,不支持再修改该库或该表的存储位置。
  • 数据库中单表的存储位置不受该库存储位置的影响。创建单表时若未指定存储位置,则会被随机放置在一个存储节点上,且后续该PolarDB-X实例上创建的所有未指定存储位置的单表,均会被放置在该存储节点上。

单库单表

建一张单库单表,不做任何拆分。


CREATE TABLE single_tbl(
 id bigint not null auto_increment, 
 name varchar(30), 
 primary key(id)
);

查看逻辑表的节点拓扑,可以看出只在0库创建了一张单库单表的逻辑表。


mysql> show topology from single_tbl;
+------+------------------------------------------------------------------+------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME |
+------+------------------------------------------------------------------+------------+
|    0 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | single_tbl |
+------+------------------------------------------------------------------+------------+
1 row in set (0.01 sec)

假设已有一个PolarDB-X实例,该实例上已有polardbx-storage-0-masterpolardbx-storage-1-master两个存储节点,其中polardbx-storage-1-master上已创建了一个数据库db1。现需要在db1库中创建一张表,并指定其的存储位置为polardbx-storage-0-master,语法如下:


mysql>CREATE TABLE tb1 (id int) LOCALITY='dn=polardbx-storage-0-master';

创建成功后,您可以通过如下语句查看该表的拓扑结构:


mysql>SHOW TOPOLOGY FROM tb1;

返回结果如下:


+----+------------------+------------+
| ID | GROUP_NAME       | TABLE_NAME | 
+----+------------------+------------+
| 0  | DB1_000000_GROUP | tb1        |
+----+------------------+------------+
1 row in set


说明 以上返回结果表示tb1表位于DB1_000000_GROUP分库。

分库不分表

假设已经建好的分库数为8,建一张表,只分库不分表,分库方式为根据ID列哈希。


CREATE TABLE multi_db_single_tbl(
  id bigint not null auto_increment, 
  name varchar(30), 
  primary key(id)
) dbpartition by hash(id);

查看该逻辑表的节点拓扑,可以看出在每个分库都创建了1张分表,既只做了分库。


mysql> show topology from multi_db_single_tbl;
+------+------------------------------------------------------------------+---------------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME          |
+------+------------------------------------------------------------------+---------------------+
|    0 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | multi_db_single_tbl |
|    1 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | multi_db_single_tbl |
|    2 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0002_RDS | multi_db_single_tbl |
|    3 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0003_RDS | multi_db_single_tbl |
|    4 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0004_RDS | multi_db_single_tbl |
|    5 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0005_RDS | multi_db_single_tbl |
|    6 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0006_RDS | multi_db_single_tbl |
|    7 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | multi_db_single_tbl |
+------+------------------------------------------------------------------+---------------------+
8 rows in set (0.01 sec)

分库分表

您可以使用如下拆分方式进行分库分表:


说明 以下示例均假设已经建好的分库数为8。

使用哈希函数做拆分

建一张表,既分库又分表,每个库含有3张物理表,分库拆分方式为按照ID列进行哈希,分表拆分方式为按照bid列进行哈希。您可以先根据ID列的值进行哈希运算,将表中数据分布在多个子库中,每个子库中的数据再根据bid列值的哈希运算结果分布在3个物理表中。


CREATE TABLE multi_db_multi_tbl(
 id bigint not null auto_increment, 
 bid int, 
 name varchar(30), 
 primary key(id)
) dbpartition by hash(id) tbpartition by hash(bid) tbpartitions 3;

查看该逻辑表的节点拓扑,可以看出在每个分库都创建了3张分表。


mysql> show topology from multi_db_multi_tbl;
+------+------------------------------------------------------------------+-----------------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME            |
+------+------------------------------------------------------------------+-----------------------+
|    0 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | multi_db_multi_tbl_00 |
|    1 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | multi_db_multi_tbl_01 |
|    2 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | multi_db_multi_tbl_02 |
|    3 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | multi_db_multi_tbl_03 |
|    4 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | multi_db_multi_tbl_04 |
|    5 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | multi_db_multi_tbl_05 |
|    6 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0002_RDS | multi_db_multi_tbl_06 |
|    7 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0002_RDS | multi_db_multi_tbl_07 |
|    8 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0002_RDS | multi_db_multi_tbl_08 |
|    9 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0003_RDS | multi_db_multi_tbl_09 |
|   10 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0003_RDS | multi_db_multi_tbl_10 |
|   11 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0003_RDS | multi_db_multi_tbl_11 |
|   12 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0004_RDS | multi_db_multi_tbl_12 |
|   13 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0004_RDS | multi_db_multi_tbl_13 |
|   14 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0004_RDS | multi_db_multi_tbl_14 |
|   15 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0005_RDS | multi_db_multi_tbl_15 |
|   16 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0005_RDS | multi_db_multi_tbl_16 |
|   17 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0005_RDS | multi_db_multi_tbl_17 |
|   18 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0006_RDS | multi_db_multi_tbl_18 |
|   19 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0006_RDS | multi_db_multi_tbl_19 |
|   20 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0006_RDS | multi_db_multi_tbl_20 |
|   21 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | multi_db_multi_tbl_21 |
|   22 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | multi_db_multi_tbl_22 |
|   23 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | multi_db_multi_tbl_23 |
+------+------------------------------------------------------------------+-----------------------+
24 rows in set (0.01 sec)

查看该逻辑表的拆分规则,可以看出分库分表的拆分方式均为哈希,分库的拆分键为ID,分表的拆分键为bid。


mysql> show rule from multi_db_multi_tbl;
+------+--------------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
| ID   | TABLE_NAME         | BROADCAST | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT |
+------+--------------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
|    0 | multi_db_multi_tbl |         0 | id               | hash                | 8                  | bid              | hash                | 3                  |
+------+--------------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
1 row in set (0.01 sec)

使用双字段哈希函数做拆分

  • 使用要求拆分键的类型必须是字符类型或数字类型。
  • 路由方式根据任一拆分键后N位计算哈希值,以哈希方式完成路由计算。N为函数第三个参数。例如RANGE_HASH(COL1, COL2, N),计算时会优先选择COL1,截取其后N位进行计算。COL1不存在时按COL2计算。
  • 适用场景适合于需要有两个拆分键,并且仅使用其中一个拆分键值进行查询时的场景。假设用户的里已经分了8个物理库, 现业务有如下的场景:
  • 一个业务想按买家ID和订单ID对订单表进行分库。
  • 查询时条件仅有买家ID或订单ID。

此时可使用以下DDL对订单表进行构建:


create table test_order_tb (
 id bigint not null auto_increment,
 seller_id varchar(30) DEFAULT NULL,
 order_id varchar(30) DEFAULT NULL,
 buyer_id varchar(30) DEFAULT NULL,
 create_time datetime DEFAULT NULL,
 primary key(id)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 dbpartition by RANGE_HASH(buyer_id, order_id, 10) tbpartition by RANGE_HASH(buyer_id, order_id, 10) tbpartitions 3;


说明

  • 两个拆分键皆不能修改。
  • 插入数据时如果发现两个拆分键指向不同的分库或分表时,插入会失败。

使用日期做拆分

除了可以使用哈希函数做拆分算法,您还可以使用日期函数MMDDWEEKMMDD来作为分表的拆分算法,具体步骤请参见如下示例。

建一张表,既分库又分表,分库方式为根据userId列哈希,分表方式为根据actionDate列,按照一周七天来拆分(WEEK(actionDate)计算的是DAY_OF_WEEK)。

比如actionDate列的值是2017-02-27,这天是星期一,WEEK(actionDate)算出的值是2,该条记录就会被存储到2(2 % 7 = 2)这张分表(位于某个分库,具体的表名是 user_log_2);比如actionDate列的值是2017-02-26,这天是星期天,WEEK(actionDate)算出的值是1,该条记录就会被存储到1(1 % 7 = 1)这张分表(位于某个分库,具体的表名是 user_log_1)。


CREATE TABLE user_log(
 userId int, 
 name varchar(30), 
 operation varchar(30), 
 actionDate DATE
) dbpartition by hash(userId) tbpartition by WEEK(actionDate) tbpartitions 7;

查看该逻辑表的节点拓扑,可以看出在每个分库都创建了7张分表(一周7天)。


mysql> show topology from user_log;
+------+------------------------------------------------------------------+------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME |
+------+------------------------------------------------------------------+------------+
|    0 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log_0 |
|    1 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log_1 |
|    2 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log_2 |
|    3 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log_3 |
|    4 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log_4 |
|    5 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log_5 |
|    6 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log_6 |
|    7 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log_0 |
|    8 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log_1 |
|    9 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log_2 |
|   10 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log_3 |
|   11 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log_4 |
|   12 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log_5 |
|   13 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log_6 |
...
|   49 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log_0 |
|   50 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log_1 |
|   51 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log_2 |
|   52 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log_3 |
|   53 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log_4 |
|   54 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log_5 |
|   55 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log_6 |
+------+------------------------------------------------------------------+------------+
56 rows in set (0.01 sec)


说明 由于返回结果较长,这里用...做了省略处理。

查看该逻辑表的拆分规则,可以看出分库的拆分方式为哈希,分库的拆分键为userId,分表的拆分方式为按照时间函数WEEK进行拆分,分表的拆分键为actionDate


mysql> show rule from user_log;
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
| ID   | TABLE_NAME | BROADCAST | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
|    0 | user_log   |         0 | userId           | hash                | 8                  | actionDate       | week                | 7                  |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
1 row in set (0.00 sec)

查看给定分库键和分表键参数时, SQL被路由到哪个物理分库和该物理分库下的哪张物理表。

建一张表,既分库又分表,分库方式为根据userId列哈希,分表方式为根据actionDate列,按照一年12个月进行拆分(MM(actionDate)计算的是MONTH_OF_YEAR)。

比如actionDate列的值是2017-02-27,MM(actionDate)算出的值是02,该条记录就会被存储到02(02 % 12 = 02)这张分表(位于某个分库,具体的表名是 user_log_02)。比如actionDate列的值是2016-12-27,MM(actionDate)算出的值是12,该条记录就会被存储到00(12 % 12 = 00)这张分表(位于某个分库,具体的表名是 user_log_00)。


CREATE TABLE user_log2(
 userId int, 
 name varchar(30), 
 operation varchar(30), 
 actionDate DATE
) dbpartition by hash(userId) tbpartition by MM(actionDate) tbpartitions 12;

查看该逻辑表的节点拓扑,可以看出在每个分库都创建了12张分表(1年有12个月)。


mysql> show topology from user_log2;
+------+------------------------------------------------------------------+--------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME   |
+------+------------------------------------------------------------------+--------------+
|    0 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_00 |
|    1 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_01 |
|    2 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_02 |
|    3 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_03 |
|    4 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_04 |
|    5 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_05 |
|    6 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_06 |
|    7 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_07 |
|    8 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_08 |
|    9 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_09 |
|   10 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_10 |
|   11 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log2_11 |
|   12 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_00 |
|   13 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_01 |
|   14 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_02 |
|   15 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_03 |
|   16 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_04 |
|   17 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_05 |
|   18 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_06 |
|   19 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_07 |
|   20 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_08 |
|   21 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_09 |
|   22 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_10 |
|   23 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0001_RDS | user_log2_11 |
...
|   84 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_00 |
|   85 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_01 |
|   86 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_02 |
|   87 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_03 |
|   88 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_04 |
|   89 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_05 |
|   90 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_06 |
|   91 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_07 |
|   92 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_08 |
|   93 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_09 |
|   94 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_10 |
|   95 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log2_11 |
+------+------------------------------------------------------------------+--------------+
96 rows in set (0.02 sec)


说明 由于返回结果较长,这里用...做了省略处理。

查看该逻辑表的拆分规则,可以看出分库的拆分方式为哈希,分库的拆分键为userId,分表的拆分方式为按照时间函数MM进行拆分,分表的拆分键为actionDate


mysql> show rule from user_log2;
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
| ID   | TABLE_NAME | BROADCAST | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
|    0 | user_log2  |         0 | userId           | hash                | 8                  | actionDate       | mm                  | 12                 |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
1 row in set (0.00 sec)

建一张表,既分库又分表,分库方式为根据userId列哈希,分表方式为按照一个月31天进行拆分(函数DD(actionDate)计算的是DAY_OF_MONTH)。

比如actionDate列的值是2017-02-27,DD(actionDate)算出的值是27,该条记录就会被存储到27(27 % 31 = 27)这张分表(位于某个分库,具体的表名是user_log_27)。


CREATE TABLE user_log3(
 userId int, 
 name varchar(30), 
 operation varchar(30), 
 actionDate DATE
) dbpartition by hash(userId) tbpartition by DD(actionDate) tbpartitions 31;

查看该逻辑表的节点拓扑,可以看出在每个分库都创建了31张分表(按每个月有31天处理)。


mysql> show topology from user_log3;
+------+------------------------------------------------------------------+--------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME   |
+------+------------------------------------------------------------------+--------------+
|    0 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_00 |
|    1 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_01 |
|    2 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_02 |
|    3 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_03 |
|    4 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_04 |
|    5 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_05 |
|    6 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_06 |
|    7 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_07 |
|    8 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_08 |
|    9 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_09 |
|   10 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_10 |
|   11 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_11 |
|   12 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_12 |
|   13 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_13 |
|   14 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_14 |
|   15 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_15 |
|   16 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_16 |
|   17 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_17 |
|   18 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_18 |
|   19 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_19 |
|   20 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_20 |
|   21 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_21 |
|   22 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_22 |
|   23 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_23 |
|   24 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_24 |
|   25 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_25 |
|   26 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_26 |
|   27 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_27 |
|   28 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_28 |
|   29 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_29 |
|   30 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log3_30 |
...
|  237 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_20 |
|  238 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_21 |
|  239 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_22 |
|  240 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_23 |
|  241 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_24 |
|  242 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_25 |
|  243 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_26 |
|  244 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_27 |
|  245 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_28 |
|  246 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_29 |
|  247 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log3_30 |
+------+------------------------------------------------------------------+--------------+
248 rows in set (0.01 sec)


说明 由于返回的结果较长,这里用...做了省略处理。

查看该逻辑表的拆分规则,可以看出分库的拆分方式为哈希,分库的拆分键为userId,分表的拆分方式为按照时间函数DD进行拆分,分表的拆分键为actionDate


mysql> show rule from user_log3;
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
| ID   | TABLE_NAME | BROADCAST | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
|    0 | user_log3  |         0 | userId           | hash                | 8                  | actionDate       | dd                  | 31                 |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
1 row in set (0.01 sec)

建一张表,既分库又分表,分库方式为根据userId列哈希,分表方式为按照一年365天进行拆分,路由到365张物理表(MMDD(actionDate) tbpartitions 365计算的是DAY_OF_YEAR % 365

比如actionDate列的值是2017-02-27,MMDD(actionDate)算出的值是58,该条记录就会被存储到58这张分表(位于某个分库,具体的表名是user_log_58)。


CREATE TABLE user_log4(
 userId int, 
 name varchar(30), 
 operation varchar(30), 
 actionDate DATE
) dbpartition by hash(userId) tbpartition by MMDD(actionDate) tbpartitions 365;

查看该逻辑表的节点拓扑,可以看出在每个分库都创建了365张分表(按每年有365天处理)。


mysql> show topology from user_log4;
+------+------------------------------------------------------------------+---------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME   |
+------+------------------------------------------------------------------+---------------+
...
| 2896 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_341 |
| 2897 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_342 |
| 2898 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_343 |
| 2899 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_344 |
| 2900 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_345 |
| 2901 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_346 |
| 2902 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_347 |
| 2903 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_348 |
| 2904 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_349 |
| 2905 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_350 |
| 2906 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_351 |
| 2907 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_352 |
| 2908 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_353 |
| 2909 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_354 |
| 2910 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_355 |
| 2911 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_356 |
| 2912 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_357 |
| 2913 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_358 |
| 2914 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_359 |
| 2915 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_360 |
| 2916 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_361 |
| 2917 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_362 |
| 2918 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_363 |
| 2919 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log4_364 |
+------+------------------------------------------------------------------+---------------+
2920 rows in set (0.07 sec)


说明 由于返回的结果较长,这里用...做了省略处理。

查看该逻辑表的拆分规则,可以看出分库的拆分方式为哈希,分库的拆分键为userId,分表的拆分方式为按照时间函数MMDD进行拆分,分表的拆分键为actionDate


mysql> show rule from user_log4;
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
| ID   | TABLE_NAME | BROADCAST | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
|    0 | user_log4  |         0 | userId           | hash                | 8                  | actionDate       | mmdd                | 365                |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
1 row in set (0.02 sec)

建一张表,既分库又分表,分库方式为根据userId列哈希,分表方式为按照一年365天进行拆分,路由到10张物理表(MMDD(actionDate) tbpartitions 10计算的是DAY_OF_YEAR % 10


CREATE TABLE user_log5(
 userId int, 
 name varchar(30), 
 operation varchar(30), 
 actionDate DATE
) dbpartition by hash(userId) tbpartition by MMDD(actionDate) tbpartitions 10;

查看该逻辑表的节点拓扑,可以看出在每个分库都创建了10张分表(按照一年365天进行拆分,路由到10张物理表)。


mysql> show topology from user_log5;
+------+------------------------------------------------------------------+--------------+
| ID   | GROUP_NAME                                                       | TABLE_NAME   |
+------+------------------------------------------------------------------+--------------+
|    0 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_00 |
|    1 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_01 |
|    2 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_02 |
|    3 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_03 |
|    4 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_04 |
|    5 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_05 |
|    6 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_06 |
|    7 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_07 |
|    8 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_08 |
|    9 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0000_RDS | user_log5_09 |
...
|   70 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_00 |
|   71 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_01 |
|   72 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_02 |
|   73 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_03 |
|   74 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_04 |
|   75 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_05 |
|   76 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_06 |
|   77 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_07 |
|   78 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_08 |
|   79 | SANGUAN_TEST_123_1488766060743ACTJSANGUAN_TEST_123_WVVP_0007_RDS | user_log5_09 |
+------+------------------------------------------------------------------+--------------+
80 rows in set (0.02 sec)


说明 由于返回的结果较长,这里用...做了省略处理。

查看该逻辑表的拆分规则,可以看出分库的拆分方式为哈希,分库的拆分键为userId,分表的拆分方式为按照时间函数 MMDD进行拆分,路由到10张物理表,分表的拆分键为actionDate


mysql> show rule from user_log5;
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
| ID   | TABLE_NAME | BROADCAST | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
|    0 | user_log5  |         0 | userId           | hash                | 8                  | actionDate       | mmdd                | 10                 |
+------+------------+-----------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+
1 row in set (0.01 sec)

广播表

子句BROADCAST用来指定创建广播表。广播表是指将这个表复制到每个分库上,在分库上通过同步机制实现数据一致,有秒级延迟。这样做的好处是可以将JOIN操作下推到底层的RDS(MySQL),来避免跨库JOIN。关于如何使用广播表来做SQL优化,详情请参见SQL调优指南


CREATE TABLE brd_tbl(
  id bigint not null auto_increment, 
  name varchar(30), 
  primary key(id)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 BROADCAST;

其他MySQL建表属性

您在分库分表的同时还可以指定其他的MySQL建表属性,例如:


CREATE TABLE multi_db_multi_tbl(
  id bigint not null auto_increment, 
  name varchar(30), 
  primary key(id)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 dbpartition by hash(id) tbpartition by hash(id) tbpartitions 3;

全局二级索引

本小节介绍如何在建表时定义全局二级索引:


说明 以下示例均假设已经建好的分库数为8。

定义全局二级索引

示例


CREATE TABLE t_order (
 `id` bigint(11) NOT NULL AUTO_INCREMENT,
 `order_id` varchar(20) DEFAULT NULL,
 `buyer_id` varchar(20) DEFAULT NULL,
 `seller_id` varchar(20) DEFAULT NULL,
 `order_snapshot` longtext DEFAULT NULL,
 `order_detail` longtext DEFAULT NULL,
 PRIMARY KEY (`id`),
 GLOBAL INDEX `g_i_seller`(`seller_id`) dbpartition by hash(`seller_id`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 dbpartition by hash(`order_id`);

其中:

  • 主表:t_order只分库不分表,分库的拆分方式为按照order_id列进行哈希。
  • 索引表:g_i_seller只分库不分表,分库的拆分方式为按照seller_id列进行哈希,未指定覆盖列。
  • 索引定义子句:GLOBAL INDEX `g_i_seller`(`seller_id`) dbpartition by hash(`seller_id`)

通过SHOW INDEX查看索引信息,包含拆分键order_id上的局部索引,和seller_ididorder_id上的GSI,其中seller_id为索引表的拆分键,idorder_id为默认的覆盖列(主键和主表的拆分键)。


说明 关于GSI的限制与约定,请参见使用全局二级索引时的注意事项 ,关于SHOW INDEX详细说明,请参见SHOW INDEX


mysql> show index from t_order;
+---------+------------+-------------------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+----------+---------------+
| TABLE   | NON_UNIQUE | KEY_NAME                | SEQ_IN_INDEX | COLUMN_NAME | COLLATION | CARDINALITY | SUB_PART | PACKED | NULL | INDEX_TYPE | COMMENT  | INDEX_COMMENT |
+---------+------------+-------------------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+----------+---------------+
| t_order |          0 | PRIMARY                 |            1 | id          | A         |           0 |     NULL | NULL   |      | BTREE      |          |               |
| t_order |          1 | auto_shard_key_order_id |            1 | order_id    | A         |           0 |     NULL | NULL   | YES  | BTREE      |          |               |
| t_order |          1 | g_i_seller              |            1 | seller_id   | NULL      |           0 |     NULL | NULL   | YES  | GLOBAL     | INDEX    |               |
| t_order |          1 | g_i_seller              |            2 | id          | NULL      |           0 |     NULL | NULL   |      | GLOBAL     | COVERING |               |
| t_order |          1 | g_i_seller              |            3 | order_id    | NULL      |           0 |     NULL | NULL   | YES  | GLOBAL     | COVERING |               |
+---------+------------+-------------------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+----------+---------------+

通过SHOW GLOBAL INDEX可以单独查看GSI信息。详细说明请参见SHOW GLOBAL INDEX


mysql> show global index from t_order;
+--------+---------+------------+------------+-------------+----------------+------------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+--------+
| SCHEMA | TABLE   | NON_UNIQUE | KEY_NAME   | INDEX_NAMES | COVERING_NAMES | INDEX_TYPE | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT | STATUS |
+--------+---------+------------+------------+-------------+----------------+------------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+--------+
| d7     | t_order | 1          | g_i_seller | seller_id   | id, order_id   | NULL       | seller_id        | HASH                | 8                  |                  | NULL                | NULL               | PUBLIC |
+--------+---------+------------+------------+-------------+----------------+------------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+--------+

查看索引表的结构,索引表包含主表的主键、分库分表键和默认的覆盖列,主键列去除了AUTO_INCREMENT属性,并且去除了主表中的局部索引。


mysql> show create table g_i_seller;
+------------+-----------------------------------------------------------+
| Table      | Create Table                                              |
+------------+-----------------------------------------------------------+
| g_i_seller | CREATE TABLE `g_i_seller` (
 `id` bigint(11) NOT NULL,
 `order_id` varchar(20) DEFAULT NULL,
 `seller_id` varchar(20) DEFAULT NULL,
 PRIMARY KEY (`id`),
 KEY `auto_shard_key_seller_id` (`seller_id`) USING BTREE
) ENGINE=InnoDB DEFAULT CHARSET=utf8 dbpartition by hash(`seller_id`) |
+------------+-----------------------------------------------------------+

定义全局唯一索引


CREATE TABLE t_order (
 `id` bigint(11) NOT NULL AUTO_INCREMENT,
 `order_id` varchar(20) DEFAULT NULL,
 `buyer_id` varchar(20) DEFAULT NULL,
 `seller_id` varchar(20) DEFAULT NULL,
 `order_snapshot` longtext DEFAULT NULL,
 `order_detail` longtext DEFAULT NULL,
 PRIMARY KEY (`id`),
 UNIQUE GLOBAL INDEX `g_i_buyer`(`buyer_id`) COVERING(`seller_id`, `order_snapshot`) 
   dbpartition by hash(`buyer_id`) tbpartition by hash(`buyer_id`) tbpartitions 3
) ENGINE=InnoDB DEFAULT CHARSET=utf8 dbpartition by hash(`order_id`);

其中:

  • 主表:t_order 只分库不分表,分库的拆分方式为按照order_id列进行哈希。
  • 索引表:g_i_buyer 只分库且分表,分库和分表的拆分方式均为按照buyer_id列进行哈希,覆盖列包含 seller_idorder_snapshot
  • 索引定义子句:UNIQUE GLOBAL INDEX `g_i_buyer`(`buyer_id`) COVERING(`seller_id`, `order_snapshot`) dbpartition by hash(`buyer_id`) tbpartition by hash(`buyer_id`) tbpartitions 3

通过SHOW INDEX 查看索引信息,包含拆分键order_id上的局部索引,和buyer_ididorder_idseller_idorder_snapshot上的GSI,其中buyer_id为索引表的拆分键,idorder_id 为默认的覆盖列(主键和主表的拆分键),seller_idorder_snapshot为显示指定的覆盖列。


说明 关于GSI的限制与约定,请参见使用全局二级索引时的注意事项 ,关于SHOW INDEX详细说明,请参见SHOW INDEX


mysql> show index from t_order;
+--------------+------------+-------------------------+--------------+----------------+-----------+-------------+----------+--------+------+------------+----------+---------------+
| TABLE        | NON_UNIQUE | KEY_NAME                | SEQ_IN_INDEX | COLUMN_NAME    | COLLATION | CARDINALITY | SUB_PART | PACKED | NULL | INDEX_TYPE | COMMENT  | INDEX_COMMENT |
+--------------+------------+-------------------------+--------------+----------------+-----------+-------------+----------+--------+------+------------+----------+---------------+
| t_order_dthb |          0 | PRIMARY                 |            1 | id             | A         |           0 |     NULL | NULL   |      | BTREE      |          |               |
| t_order_dthb |          1 | auto_shard_key_order_id |            1 | order_id       | A         |           0 |     NULL | NULL   | YES  | BTREE      |          |               |
| t_order      |          0 | g_i_buyer               |            1 | buyer_id       | NULL      |           0 |     NULL | NULL   | YES  | GLOBAL     | INDEX    |               |
| t_order      |          1 | g_i_buyer               |            2 | id             | NULL      |           0 |     NULL | NULL   |      | GLOBAL     | COVERING |               |
| t_order      |          1 | g_i_buyer               |            3 | order_id       | NULL      |           0 |     NULL | NULL   | YES  | GLOBAL     | COVERING |               |
| t_order      |          1 | g_i_buyer               |            4 | seller_id      | NULL      |           0 |     NULL | NULL   | YES  | GLOBAL     | COVERING |               |
| t_order      |          1 | g_i_buyer               |            5 | order_snapshot | NULL      |           0 |     NULL | NULL   | YES  | GLOBAL     | COVERING |               |
+--------------+------------+-------------------------+--------------+----------------+-----------+-------------+----------+--------+------+------------+----------+---------------+

通过SHOW GLOBAL INDEX可以单独查看GSI信息。详细说明请参见SHOW GLOBAL INDEX


mysql> show global index from t_order;
+--------+---------+------------+-----------+-------------+-----------------------------------------+------------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+--------+
| SCHEMA | TABLE   | NON_UNIQUE | KEY_NAME  | INDEX_NAMES | COVERING_NAMES                          | INDEX_TYPE | DB_PARTITION_KEY | DB_PARTITION_POLICY | DB_PARTITION_COUNT | TB_PARTITION_KEY | TB_PARTITION_POLICY | TB_PARTITION_COUNT | STATUS |
+--------+---------+------------+-----------+-------------+-----------------------------------------+------------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+--------+
| d7     | t_order | 0          | g_i_buyer | buyer_id    | id, order_id, seller_id, order_snapshot | NULL       | buyer_id         | HASH                | 8                  | buyer_id         | HASH                | 3                  | PUBLIC |
+--------+---------+------------+-----------+-------------+-----------------------------------------+------------+------------------+---------------------+--------------------+------------------+---------------------+--------------------+--------+

查看索引表的结构,索引表包含主表的主键、分库分表键、默认覆盖列和GSI定义中指定的覆盖列,主键列去除了AUTO_INCREMENT属性,并且去除了主表中局部索引,全局唯一索引默认会创建一份数据表来实现全局的唯一性支持。


mysql> show create table g_i_buyer;
+-----------+--------------------------------------------------------------------------------------------------------+
| Table     | Create Table                                                                                           |
+-----------+--------------------------------------------------------------------------------------------------------+
| g_i_buyer | CREATE TABLE `g_i_buyer` (
  `id` bigint(11) NOT NULL,
  `order_id` varchar(20) DEFAULT NULL,
  `buyer_id` varchar(20) DEFAULT NULL,
  `seller_id` varchar(20) DEFAULT NULL,
  `order_snapshot` longtext,
  PRIMARY KEY (`id`),
  UNIQUE KEY `auto_shard_key_buyer_id` (`buyer_id`) USING BTREE
) ENGINE=InnoDB DEFAULT CHARSET=utf8 dbpartition by hash(`buyer_id`) tbpartition by hash(`buyer_id`) tbpartitions 3 |
+-----------+-----------------------------------------------------------------------------------------------------
相关实践学习
快速体验PolarDB开源数据库
本实验环境已内置PostgreSQL数据库以及PolarDB开源数据库:PolarDB PostgreSQL版和PolarDB分布式版,支持一键拉起使用,方便各位开发者学习使用。
相关文章
|
Java iOS开发 MacOS
学妹:你的 IDEA 主题怎么这么好看?
哈喽,小伙伴们好呀。我是狗哥,今天不聊技术,聊聊我们的吃饭工具。经常有小伙伴问我:“你的 IDEA 主题怎么这么好看,能推荐一下不?”。问的人太多了,才索性写了这篇文章。 在这篇文章中,我精选了几个比较适合 Java 编码的 IDEA 主题供小伙伴们选择。另外,我自己用的是 One Dark theme 这款。
学妹:你的 IDEA 主题怎么这么好看?
|
SQL 存储 运维
阿里云分布式关系型数据库服务 DRDS
DRDS 是阿里巴巴集团自主研发的分布式数据库中间件产品,专注于解决单机关系型数据库扩展性问题,具备轻量(无状态)、灵活、稳定、高效等特性,稳定运行11年,经历历届双十一核心交易业务和各类行业业务的考验
11592 0
|
存储 SQL 算法
【笔记】开发指南—DDL语句—分库分表语法—CREATE TABLE
本文主要介绍使用DDL语句进行建表的语法、子句、参数和基本方式。
124 0
|
SQL 关系型数据库 MySQL
开发指南—DDL语句—分库分表语法—DROP TABLE
您可以使用DROP TABLE语法删除指定表。
|
SQL 关系型数据库 MySQL
开发指南—DDL语句—分库分表语法—DROP INDEX
本文介绍了如何删除局部索引和全局二级索引
|
存储 SQL 算法
【笔记】开发指南—DDL语句—分库分表语法—CREATE TABLE
本文主要介绍使用DDL语句进行建表的语法、子句、参数和基本方式。
145 0
|
存储 SQL 算法
开发指南—DDL语句—分库分表语法—CREATE TABLE
本文主要介绍使用DDL语句进行建表的语法、子句、参数和基本方式。
206 0
|
SQL 关系型数据库 MySQL
开发指南—DDL语句—分库分表语法—ALTER TABLE
您可以通过ALTER TABLE语法改变表的结构,如增加列、增加索引、修改数据定义等
117 0
|
SQL 关系型数据库 MySQL
开发指南—DDL语句—分库分表语法—CREATE INDEX
PolarDB-X支持创建局部索引和全局二级索引 (Global Secondary Index, GSI) ,同时支持删除这两种索引。
107 0
|
SQL 索引
开发指南—DDL语句—分库分表语法—RENAME TABLE
本文介绍了对表做重命名的方法
105 0
下一篇
无影云桌面