一、简介
在数据库日渐庞大的今天，为了方便对数据库数据的管理，比如按时间，按地区去统计一些数据时，基数过于庞大，多有不便。很多商业数据库都提供分区的概念，按不同的维度去存放数据，便于后期的管理，PostgreSQL也不例外。
PostgresSQL分区的意思是把逻辑上的一个大表分割成物理上的几块儿。分区不仅能带来访问速度的提升，关键的是，它能带来管理和维护上的方便。
分区的具体好处是：
- 某些类型的查询性能可以得到极大提升。
- 更新的性能也可以得到提升，因为表的每块的索引要比在整个数据集上的索引要小。如果索引不能全部放在内存里，那么在索引上的读和写都会产生更多的磁盘访问。
- 批量删除可以用简单的删除某个分区来实现。
- 可以将很少用的数据移动到便宜的、转速慢的存储介质上。
在PG里表分区是通过表继承来实现的，一般都是建立一个主表，里面是空，然后每个分区都去继承它。无论何时，都应保证主表里面是空的。
小表分区不实际，表在多大情况下才考虑分区呢？PostgresSQL官方给出的建议是：当表本身大小超过了机器物理内存的实际大小时（the size of the table should exceed the physical memory of the database server），可以考虑分区。
PG目前（9.2.2）仅支持范围分区和列表分区，尚未支持散列分区。
二、环境
系统环境：CentOS release 6.3 (Final)
PostgreSQL版本：PostgreSQL 9.2.2 on x86_64-unknown-linux-gnu, compiled by gcc (GCC) 4.4.6 20120305 (Red Hat 4.4.6-4), 64-bit
三、实现分区
3.1 创建主表
david=# create table tbl_partition ( david(# id integer, david(# name varchar(20), david(# gender boolean, david(# join_date date, david(# dept char(4)); CREATE TABLE david=#
3.2 创建分区表
david=# create table tbl_partition_201211 ( check ( join_date >= DATE '2012-11-01' AND join_date < DATE '2012-12-01' ) ) INHERITS (tbl_partition); CREATE TABLE david=# create table tbl_partition_201212 ( check ( join_date >= DATE '2012-12-01' AND join_date < DATE '2013-01-01' ) ) INHERITS (tbl_partition); CREATE TABLE david=# create table tbl_partition_201301 ( check ( join_date >= DATE '2013-01-01' AND join_date < DATE '2013-02-01' ) ) INHERITS (tbl_partition); CREATE TABLE david=# create table tbl_partition_201302 ( check ( join_date >= DATE '2013-02-01' AND join_date < DATE '2013-03-01' ) ) INHERITS (tbl_partition); CREATE TABLE david=# create table tbl_partition_201303 ( check ( join_date >= DATE '2013-03-01' AND join_date < DATE '2013-04-01' ) ) INHERITS (tbl_partition); CREATE TABLE david=# create table tbl_partition_201304 ( check ( join_date >= DATE '2013-04-01' AND join_date < DATE '2013-05-01' ) ) INHERITS (tbl_partition); CREATE TABLE david=# create table tbl_partition_201305 ( check ( join_date >= DATE '2013-05-01' AND join_date < DATE '2013-06-01' ) ) INHERITS (tbl_partition); CREATE TABLE david=#
3.3 分区键上建索引
david=# create index tbl_partition_201211_joindate on tbl_partition_201211 (join_date); CREATE INDEX david=# create index tbl_partition_201212_joindate on tbl_partition_201212 (join_date); CREATE INDEX david=# create index tbl_partition_201301_joindate on tbl_partition_201301 (join_date); CREATE INDEX david=# create index tbl_partition_201302_joindate on tbl_partition_201302 (join_date); CREATE INDEX david=# create index tbl_partition_201303_joindate on tbl_partition_201303 (join_date); CREATE INDEX david=# create index tbl_partition_201304_joindate on tbl_partition_201304 (join_date); CREATE INDEX david=# create index tbl_partition_201305_joindate on tbl_partition_201305 (join_date); CREATE INDEX david=#
对于开发人员来说，希望数据库是透明的，只管 insert into tbl_partition。对于数据插向哪个分区，则希望由DB决定。这点，ORACLE实现了，但是PG不行，需要前期人工处理下。
3.4 创建触发器函数
david=# CREATE OR REPLACE FUNCTION tbl_partition_insert_trigger() RETURNS TRIGGER AS $$ BEGIN IF ( NEW.join_date >= DATE '2012-11-01' AND NEW.join_date < DATE '2012-12-01' ) THEN INSERT INTO tbl_partition_201211 VALUES (NEW.*); ELSIF ( NEW.join_date >= DATE '2012-12-01' AND NEW.join_date < DATE '2013-01-01' ) THEN INSERT INTO tbl_partition_201212 VALUES (NEW.*); ELSIF ( NEW.join_date >= DATE '2013-01-01' AND NEW.join_date < DATE '2013-02-01' ) THEN INSERT INTO tbl_partition_201301 VALUES (NEW.*); ELSIF ( NEW.join_date >= DATE '2013-02-01' AND NEW.join_date < DATE '2013-03-01' ) THEN INSERT INTO tbl_partition_201302 VALUES (NEW.*); ELSIF ( NEW.join_date >= DATE '2013-03-01' AND NEW.join_date < DATE '2013-04-01' ) THEN INSERT INTO tbl_partition_201303 VALUES (NEW.*); ELSIF ( NEW.join_date >= DATE '2013-04-01' AND NEW.join_date < DATE '2013-05-01' ) THEN INSERT INTO tbl_partition_201304 VALUES (NEW.*); ELSIF ( NEW.join_date >= DATE '2013-05-01' AND NEW.join_date < DATE '2013-06-01' ) THEN INSERT INTO tbl_partition_201305 VALUES (NEW.*); ELSE RAISE EXCEPTION 'Date out of range. Fix the tbl_partition_insert_trigger() function!'; END IF; RETURN NULL; END; $$ LANGUAGE plpgsql; CREATE FUNCTION david=#
说明：如果不想丢失数据，上面的ELSE 条件可以改成 INSERT INTO tbl_partition_error_join_date VALUES (NEW.*); 同时需要创建一张结构和tbl_partition 一样的表tbl_partition_error_join_date，这样，错误的join_date 数据就可以插入到这张表中而不是报错了。
3.5 创建触发器
david=# CREATE TRIGGER insert_tbl_partition_trigger david-# BEFORE INSERT ON tbl_partition david-# FOR EACH ROW EXECUTE PROCEDURE tbl_partition_insert_trigger(); CREATE TRIGGER david=#
四、查看表
4.1 查看所有表
david=# \dt List of relations Schema | Name | Type | Owner --------+----------------------+-------+---------- public | tbl_partition | table | postgres public | tbl_partition_201211 | table | postgres public | tbl_partition_201212 | table | postgres public | tbl_partition_201301 | table | postgres public | tbl_partition_201302 | table | postgres public | tbl_partition_201303 | table | postgres public | tbl_partition_201304 | table | postgres public | tbl_partition_201305 | table | postgres (8 rows) david=#
4.2 查看主表
david=# \d tbl_partition Table "public.tbl_partition" Column | Type | Modifiers -----------+-----------------------+----------- id | integer | name | character varying(20) | gender | boolean | join_date | date | dept | character(4) | Triggers: insert_tbl_partition_trigger BEFORE INSERT ON tbl_partition FOR EACH ROW EXECUTE PROCEDURE tbl_partition_insert_trigger() Number of child tables: 7 (Use \d+ to list them.) david=#
4.3 查看分区表
david=# \d tbl_partition_201304 Table "public.tbl_partition_201304" Column | Type | Modifiers -----------+-----------------------+----------- id | integer | name | character varying(20) | gender | boolean | join_date | date | dept | character(4) | Indexes: "tbl_partition_201304_joindate" btree (join_date) Check constraints: "tbl_partition_201304_join_date_check" CHECK (join_date >= '2013-04-01'::date AND join_date < '2013-05-01'::date) Inherits: tbl_partition david=#
五、测试
5.1 插入数据
david=# insert into tbl_partition values (1, 'David', '1', '2013-01-10', 'TS'); INSERT 0 0 david=# insert into tbl_partition values (2, 'Sandy', '0', '2013-02-10', 'TS'); INSERT 0 0 david=# insert into tbl_partition values (3, 'Eagle', '1', '2012-11-01', 'TS'); INSERT 0 0 david=# insert into tbl_partition values (4, 'Miles', '1', '2012-12-15', 'SD'); INSERT 0 0 david=# insert into tbl_partition values (5, 'Simon', '1', '2012-12-10', 'SD'); INSERT 0 0 david=# insert into tbl_partition values (6, 'Rock', '1', '2012-11-10', 'SD'); INSERT 0 0 david=# insert into tbl_partition values (7, 'Peter', '1', '2013-01-11', 'SD'); INSERT 0 0 david=# insert into tbl_partition values (8, 'Sally', '0', '2013-03-10', 'BCSC'); INSERT 0 0 david=# insert into tbl_partition values (9, 'Carrie', '0', '2013-04-02', 'BCSC'); INSERT 0 0 david=# insert into tbl_partition values (10, 'Lee', '1', '2013-01-05', 'BMC'); INSERT 0 0 david=# insert into tbl_partition values (11, 'Nicole', '0', '2012-11-10', 'PROJ'); INSERT 0 0 david=# insert into tbl_partition values (12, 'Renee', '0', '2013-01-10', 'TS'); INSERT 0 0 david=#
5.2 查看主表数据
david=# select * from tbl_partition; id | name | gender | join_date | dept ----+--------+--------+------------+------ 3 | Eagle | t | 2012-11-01 | TS 6 | Rock | t | 2012-11-10 | SD 11 | Nicole | f | 2012-11-10 | PROJ 4 | Miles | t | 2012-12-15 | SD 5 | Simon | t | 2012-12-10 | SD 1 | David | t | 2013-01-10 | TS 7 | Peter | t | 2013-01-11 | SD 10 | Lee | t | 2013-01-05 | BMC 12 | Renee | f | 2013-01-10 | TS 2 | Sandy | f | 2013-02-10 | TS 8 | Sally | f | 2013-03-10 | BCSC 9 | Carrie | f | 2013-04-02 | BCSC (12 rows) david=#
5.3 查看分区表数据
david=# select * from tbl_partition_201301 ; id | name | gender | join_date | dept ----+-------+--------+------------+------ 1 | David | t | 2013-01-10 | TS 7 | Peter | t | 2013-01-11 | SD 10 | Lee | t | 2013-01-05 | BMC 12 | Renee | f | 2013-01-10 | TS (4 rows) david=#
六、管理分区
6.1 移除数据/分区
实现分区表之后，我们就可以很容易地移除不再使用的旧数据了，最简单的方法就是：
david=# drop table tbl_partition_201304;
这样可以快速移除大量数据，而不是逐条删除数据。
另一个推荐做法是将分区从分区表中移除，但是保留访问权限。
david=# alter table tbl_partition_201304 no inherit tbl_partition; ALTER TABLE david=#
和直接DROP 相比，该方式仅仅是使子表脱离了原有的主表，而存储在子表中的数据仍然可以得到访问，因为此时该表已经被还原成一个普通的数据表了。这样对于数据库的DBA来说，就可以在此时对该表进行必要的维护操作，如数据清理、归档等，在完成诸多例行性的操作之后，就可以考虑是直接删除该表(DROP TABLE)，还是先清空该表的数据(TRUNCATE TABLE)，之后再让该表重新继承主表。
david=# alter table tbl_partition_201304 inherit tbl_partition; ALTER TABLE david=#
6.2 增加分区
我们可以像之前那样增加一个分区
david=# create table tbl_partition_201306 ( check ( join_date >= DATE '2013-06-01' AND join_date < DATE '2013-07-01' ) ) INHERITS (tbl_partition);
david=# create index tbl_partition_201306_joindate on tbl_partition_201306 (join_date);
同时，需要修改触发器函数，将插入条件改成相应的值。
说明：创建触发器函数时，最好把插入条件写更未来一点，比如多写十年，这样以后增加新分区时就不需要重新创建触发器函数了，也可以避免一些不必要的错误。
另外，还可以如下增加新的分区：
david=# create table tbl_partition_201307 david-# (LIKE tbl_partition INCLUDING DEFAULTS INCLUDING CONSTRAINTS); CREATE TABLE david=# david=# alter table tbl_partition_201307 add constraint tbl_partition_201307_join_date_check david-# check ( join_date >= DATE '2013-07-01' AND join_date < DATE '2013-08-01' ); ALTER TABLE david=# david=# create index tbl_partition_201307_joindate on tbl_partition_201307 (join_date);
david=# copy tbl_partition_201307 from '/tmp/tbl_partition_201307.sql'; //从文件中拷贝数据，这些数据可以是事前准备的 david=# alter table tbl_partition_201307 inherit tbl_partition;
七、约束排除
约束排除（Constraint exclusion）是一种查询优化技巧，它改进了用上面方法定义的表分区的性能。
确保postgresql.conf 里的配置参数constraint_exclusion 是打开的。没有这个参数，查询不会按照需要进行优化。这里我们需要做的是确保该选项在配置文件中没有被注释掉。
如果没有约束排除，查询会扫描tbl_partition 表中的每一个分区。打开了约束排除之后，规划器将检查每个分区的约束然后再试图证明该分区不需要被扫描，因为它不能包含任何符合WHERE子句条件的数据行。如果规划器可以证明这个，它就把该分区从查询规划里排除出去。
可以使用EXPLAIN 命令显示一个规划在constraint_exclusion 关闭和打开情况下的不同：
7.1 约束排除关闭
david=# set constraint_exclusion = off; SET david=# explain select count(*) from tbl_partition where join_date >= DATE '2013-04-01'; QUERY PLAN ------------------------------------------------------------------------------------------------- Aggregate (cost=172.80..172.81 rows=1 width=0) -> Append (cost=0.00..167.62 rows=2071 width=0) -> Seq Scan on tbl_partition (cost=0.00..0.00 rows=1 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201211 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201212 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201301 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201302 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201303 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201305 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201304 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201306 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201307 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) (22 rows) david=#
从上面的查询计划中可以看出，PostgreSQL 扫描了所有分区。下面我们再看一下打开约束排除之后的查询计划：
7.2 约束排除开启
david=# set constraint_exclusion = on; SET david=# explain select count(*) from tbl_partition where join_date >= DATE '2013-04-01'; QUERY PLAN ------------------------------------------------------------------------------------------------- Aggregate (cost=76.80..76.81 rows=1 width=0) -> Append (cost=0.00..74.50 rows=921 width=0) -> Seq Scan on tbl_partition (cost=0.00..0.00 rows=1 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201305 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201304 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201306 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) -> Seq Scan on tbl_partition_201307 tbl_partition (cost=0.00..18.62 rows=230 width=0) Filter: (join_date >= '2013-04-01'::date) (12 rows) david=#
可以看到，PostgreSQL 只扫描四月份以后的分区表。
八、可选的分区方式
还可以通过建立规则的方式进行分区。
CREATE RULE insert_tbl_partition_201211 AS ON INSERT TO tbl_partition WHERE ( join_date >= DATE '2012-11-01' AND join_date < DATE '2012-12-01' ) DO INSTEAD INSERT INTO tbl_partition_201211 VALUES (NEW.*); CREATE RULE insert_tbl_partition_201212 AS ON INSERT TO tbl_partition WHERE ( join_date >= DATE '2012-12-01' AND join_date < DATE '2013-01-01' ) DO INSTEAD INSERT INTO tbl_partition_201212 VALUES (NEW.*); ... CREATE RULE insert_tbl_partition_201306 AS ON INSERT TO tbl_partition WHERE ( join_date >= DATE '2013-06-01' AND join_date < DATE '2013-07-01' ) DO INSTEAD INSERT INTO tbl_partition_201306 VALUES (NEW.*); CREATE RULE insert_tbl_partition_201307 AS ON INSERT TO tbl_partition WHERE ( join_date >= DATE '2013-07-01' AND join_date < DATE '2013-08-01' ) DO INSTEAD INSERT INTO tbl_partition_201307 VALUES (NEW.*); CREATE RULE insert_tbl_partition_error_join_date AS ON INSERT TO tbl_partition WHERE ( join_date >= DATE '2013-08-01' OR join_date < DATE '2012-11-01' ) DO INSTEAD INSERT INTO tbl_partition_error_join_date VALUES (NEW.*);
九、注意事项
VACUUM 或 ANALYZE tbl_partition 只会对主表起作用，要想分析表，需要分别分析每个分区表。
十、参考资料
- PostgreSQL官方说明：http://www.postgresql.org/docs/9.2/static/ddl-partitioning.html
- ITEYE：http://diegoball.iteye.com/blog/713826
- kenyon(君羊)：http://my.oschina.net/Kenyon/blog/59455