http://blog.csdn.net/goldenhawking
最近发现大家越来越关注 PostgreSQL了。2008年以来,通过对PostgreSQL的实际使用,发现其对象-关系数据库机制对现实问题非常有帮助。在多重表继承下,对上亿条不同类别的数据条目进行按型号、按月份双层分区管理,既可在总表查阅所有条目的共有字段,也可在各类型字表查询附加字段,非常高效。下面把这种分区机制介绍如下!
实验平台:PostgreSQL 9.1
实验背景:
假设有N种数据收集设备,分别叫做 machine1, machine2...machineN, 各类收集设备从传感器上采集的数据各不相同。但是他们都包括3个共有属性:1、采集时刻 2、一个电压值 3、机器的ID。 这些机器源源不断的从各个传感器收集信息,每类机器还有各自不同的附加数据。比如,machine1有当前最大单元数、当前已使用单元数两个属性。Machine2有前端传感器的ID和取值。数据量约100万条/天,要求数据库容纳至少5年的数据。
设计原则:
由于采集的频率高,每天会有上百万条数据存入,为了考虑缩小索引的规模,提高检索效率,采用按月分区存储。由于各类机器的字段各有区别,使得我们必须设计不同的表结构, 分别存储各类数据。由于要求能够统一检索基本信息、按需检索额外信息,我们采用PostgreSQL的表继承,首先按照机器类型分类,而后各类型机器内按照月份分类。
数据库结构:
全局ID 序列:
- CREATE SEQUENCE serial_id_seq
- INCREMENT 1
- MINVALUE 1
- MAXVALUE 9223372036854775807
- START 1
- CACHE 1;
- ALTER TABLE serial_id_seq
- OWNER TO postgres;
该序列用于保持全局ID的唯一性。PostgreSQL各个继承表中的主键约束仅仅限于本表,在不想通过检查条件确保唯一的情况下,可以通过触发器手工从序列获取新的值,以及限制用户修改ID来保证唯一。基本表(爷爷表),承载了所有机器的共同属性
- CREATE TABLE base_table
- (
- id bigint NOT NULL,
- dvalue double precision,
- sample_time timestamp with time zone,
- machine_code character varying(32),
- CONSTRAINT pk_base_table_id PRIMARY KEY (id )
- )
- WITH (
- OIDS=FALSE
- );
- ALTER TABLE base_table
- OWNER TO postgres;
- CREATE INDEX idx_sample_time
- ON base_table
- USING btree
- (sample_time );
下面为机器类型1创建按类型分区子表(爸爸表)
- CREATE TABLE base_table_machine1
- (
- max_res integer,
- curr_res integer,
- CONSTRAINT pk_base_table_machine1 PRIMARY KEY (id )
- )
- INHERITS (base_table)
- WITH (
- OIDS=FALSE
- );
- ALTER TABLE base_table_machine1
- OWNER TO postgres;
- CREATE INDEX idx_base_table_machine1_sample_time
- ON base_table_machine1
- USING btree
- (sample_time );
同样,为机器2创建按类型分区子表
- CREATE TABLE base_table_machine2
- (
- manu_id character varying(16),
- manu_value character varying(16),
- CONSTRAINT pk_base_table_machine2 PRIMARY KEY (id )
- )
- INHERITS (base_table)
- WITH (
- OIDS=FALSE
- );
- ALTER TABLE base_table_machine2
- OWNER TO postgres;
- CREATE INDEX idx_base_table_machine2_sample_time
- ON base_table_machine2
- USING btree
- (sample_time );
其他机器不再赘述。创建完后,我们开始写创建按月分区表的触发器(儿子表)。按月分区会判断每次插入的数据的时刻,按照月份放到分区表中。如果分区表不存在,则自动创建。这里给出机器1、机器2 的触发器
- -- Function: on_insert_base_table_machine1()
- -- DROP FUNCTION on_insert_base_table_machine1();
- CREATE OR REPLACE FUNCTION on_insert_base_table_machine1()
- RETURNS trigger AS
- $BODY$
- DECLARE
- --Variable Hold subtable name
- str_sub_tablename varchar;
- --Variable Hold year\month info with timestamle
- str_sub_sample_time varchar;
- str_sql_cmd varchar;
- str_sub_checkval varchar;
- BEGIN
- --The triggle func will be exectued only when BEFORE INSERT
- IF TG_OP <> 'INSERT' OR TG_TABLE_NAME <>'base_table_machine1' OR TG_WHEN <> 'BEFORE' THEN
- RETURN NULL;
- END IF;
- --Generate Table Name
- str_sub_sample_time = date_part('year',NEW.sample_time)::varchar || '_' ||
- CASE WHEN date_part('month',NEW.sample_time) <10 THEN '0' ELSE '' END
- ||date_part('month',NEW.sample_time)::varchar;
- str_sub_tablename = 'machine1_' || str_sub_sample_time;
- --Check if table not created
- select * from pg_tables where schemaname = 'public' and tablename=str_sub_tablename
- into str_sql_cmd;
- IF NOT FOUND THEN
- --Create table Cmd
- str_sql_cmd = '
- CREATE TABLE '||str_sub_tablename||'
- (
- CONSTRAINT pk_'|| str_sub_tablename||' PRIMARY KEY (id ),
- CONSTRAINT chk_'|| str_sub_tablename||'
- CHECK(date_part(''year''::text, sample_time) = '||
- date_part('year',NEW.sample_time)::varchar||
- '::double precision AND
- date_part(''month''::text, sample_time) = '||
- date_part('month',NEW.sample_time)::varchar||'
- )
- )
- INHERITS (base_table_machine1)
- WITH ( OIDS=FALSE );
- ALTER TABLE '||str_sub_tablename||' OWNER TO postgres;
- CREATE INDEX idx_'|| str_sub_tablename||'_sample_time
- ON '|| str_sub_tablename||'
- USING btree (sample_time );
- ';
- EXECUTE str_sql_cmd;
- END IF;
- --insert Data
- str_sql_cmd = 'INSERT INTO '||str_sub_tablename||'
- ( id,dvalue,sample_time,machine_code,max_res,curr_res) VALUES (
- nextval(''serial_id_seq''),$1,$2,$3,$4,$5);
- ';
- EXECUTE str_sql_cmd USING
- NEW.dvalue,
- NEW.sample_time,
- NEW.machine_code,
- NEW.max_res,
- NEW.curr_res;
- --return null because main table does not really contain data
- return NULL;
- END;
- $BODY$
- LANGUAGE plpgsql VOLATILE
- COST 100;
- ALTER FUNCTION on_insert_base_table_machine1()
- OWNER TO postgres;
- -- Function: on_insert_base_table_machine2()
- -- DROP FUNCTION on_insert_base_table_machine2();
- CREATE OR REPLACE FUNCTION on_insert_base_table_machine2()
- RETURNS trigger AS
- $BODY$
- DECLARE
- --Variable Hold subtable name
- str_sub_tablename varchar;
- --Variable Hold year\month info with timestamle
- str_sub_sample_time varchar;
- str_sql_cmd varchar;
- str_sub_checkval varchar;
- BEGIN
- --The triggle func will be exectued only when BEFORE INSERT
- IF TG_OP <> 'INSERT' OR TG_TABLE_NAME <>'base_table_machine2' OR TG_WHEN <> 'BEFORE' THEN
- RETURN NULL;
- END IF;
- --Generate Table Name
- str_sub_sample_time = date_part('year',NEW.sample_time)::varchar || '_' ||
- CASE WHEN date_part('month',NEW.sample_time) <10 THEN '0' ELSE '' END
- ||date_part('month',NEW.sample_time)::varchar;
- str_sub_tablename = 'machine2_' || str_sub_sample_time;
- --Check if table not created
- select * from pg_tables where schemaname = 'public' and tablename=str_sub_tablename
- into str_sql_cmd;
- IF NOT FOUND THEN
- --Create table Cmd
- str_sql_cmd = '
- CREATE TABLE '||str_sub_tablename||'
- (
- CONSTRAINT pk_'|| str_sub_tablename||' PRIMARY KEY (id ),
- CONSTRAINT chk_'|| str_sub_tablename||'
- CHECK(date_part(''year''::text, sample_time) = '||
- date_part('year',NEW.sample_time)::varchar||
- '::double precision AND
- date_part(''month''::text, sample_time) = '||
- date_part('month',NEW.sample_time)::varchar||'
- )
- )
- INHERITS (base_table_machine2)
- WITH ( OIDS=FALSE );
- ALTER TABLE '||str_sub_tablename||' OWNER TO postgres;
- CREATE INDEX idx_'|| str_sub_tablename||'_sample_time
- ON '|| str_sub_tablename||'
- USING btree (sample_time );
- ';
- EXECUTE str_sql_cmd;
- END IF;
- --insert Data
- str_sql_cmd = 'INSERT INTO '||str_sub_tablename||'
- ( id,dvalue,sample_time,machine_code,manu_id,manu_value) VALUES (
- nextval(''serial_id_seq''),$1,$2,$3,$4,$5);
- ';
- EXECUTE str_sql_cmd USING
- NEW.dvalue,
- NEW.sample_time,
- NEW.machine_code,
- NEW.manu_id,
- NEW.manu_value;
- --return null because main table does not really contain data
- return NULL;
- END;
- $BODY$
- LANGUAGE plpgsql VOLATILE
- COST 100;
- ALTER FUNCTION on_insert_base_table_machine2()
- OWNER TO postgres;
最后,为各个爸爸表设置触发器
- CREATE TRIGGER triggle_on_insert_machine1
- BEFORE INSERT
- ON base_table_machine1
- FOR EACH ROW
- EXECUTE PROCEDURE on_insert_base_table_machine1();
- CREATE TRIGGER triggle_machine2
- BEFORE INSERT
- ON base_table_machine2
- FOR EACH ROW
- EXECUTE PROCEDURE on_insert_base_table_machine2();
到此为止,我们可以分别向各个爸爸表(按类型分区表)插入数据,而后通过爷爷表(总表)检索基本信息,通过爸爸表检索详细信息。对总表的操作会遍历反馈到所有子表,试图利用子表的索引进行查询。由于按月存储,插入工作只限于本月,所以检索历史数据效率很高。
当然了,这只是简单的实验,实际字段要比上述字段复杂很多。PostgreSQL的对象-关系数据库对解决上述问题非常有帮助,也全面的运用到我公司的各个环节,达到工业化标准的系统非常稳定,尽管设置了备份,但4年来从未真正用到。我们目前使用 16核心机架服务器,8GB内存,Ubuntu 12.04 LTS,优化配置(Postgresql.conf) 采用设置共享段shared_buffers 512MB, work_mem 32MB,维护maintenance_work_mem 512MB,checkpoint_segments = 16,获得了稳定而持久的生产力提升。
测试:
插入4条数据
- insert into base_table_machine1 (dvalue,sample_time,machine_code,max_res,curr_res) values (22.17273,'2012-06-01 11:22:11','SC3010-192.168.1.12',1,2);
- insert into base_table_machine1 (dvalue,sample_time,machine_code,max_res,curr_res) values (12.8273,'2012-07-12 10:23:01','SC3010-192.168.1.14',1,2);
- insert into base_table_machine2 (dvalue,sample_time,machine_code,manu_id,manu_value) values (4412.1928,'2011-01-21 02:08:34','PK937-192.168.1.113','TP1','E54DF');
- insert into base_table_machine2 (dvalue,sample_time,machine_code,manu_id,manu_value) values (4412.1928,'2011-12-31 04:21:31','PK937-192.168.1.112','TP2','CB67D');
看看 select 语句的结果
select * from base_table;
select * from base_table_machine2;
select * from base_table_machine1;
explain select * from base_table where sample_time >='2012-06-21 00:00:00' and sample_time <='2012-07-21 00:00:00';
自动使用索引关联到每个子表。