PostgreSQL, SQL Server 逻辑增量 (通过逻辑标记update,delete) 同步到 Greenplum, PostgreSQL-阿里云开发者社区

create table t_src (  
  id int primary key,   
  info text not null,   
  is_del boolean default null,   -- 删除标记，NULL表示未删除，非空表示已删除  
  mod_time timestamp not null default clock_timestamp()  -- 插入、删除、修改的时间戳  
);

2、创建索引，加速同步

create index idx_t_src_1 on t_src(mod_time);

3、创建触发器函数，更新、删除数据时，更新时间戳，同时修改删除标记位

当被删除的记录重新被插入时，把删除标记改成未删除。

create or replace function tg1_t_src() returns trigger as $$  
declare  
begin  
  NEW.mod_time := clock_timestamp();  
  select case when OLD.is_del is null and NEW.is_del = true then true else null end into NEW.is_del;     -- 如果以前这个ID被删除过，则插入，并将is_del重新置为未删除   
  return NEW;  
end;  
$$ language plpgsql strict;

4、创建触发器，更新时触发

create trigger tg1 before update on t_src for each row execute procedure tg1_t_src();

5、创建规则，当删除记录时，使用UPDATE代替DELETE

create rule r1 as on delete to t_src do instead update t_src set is_del=true where t_src.id=OLD.id and t_src.is_del is null;     -- 未标记为删除的记录is_del=null，标记为删除.

6、查看插入、更新、删除是否符合预期

postgres=# insert into t_src values (1, md5(random()::text)) on conflict (id) do update set info=excluded.info;  
INSERT 0 1  
postgres=# select * from t_src where id=1;  
 id |               info               | is_del |          mod_time            
----+----------------------------------+--------+----------------------------  
  1 | 56c21963342997fd8bf80a5b542abde9 |        | 2018-05-12 08:54:19.393532  
(1 row)  
  
postgres=# insert into t_src values (1, md5(random()::text)) on conflict (id) do update set info=excluded.info;  
INSERT 0 1  
postgres=# select * from t_src where id=1;  
 id |               info               | is_del |          mod_time            
----+----------------------------------+--------+----------------------------  
  1 | 5bca407559081d6cfc1154fd0f17b6a9 |        | 2018-05-12 08:54:23.465005  
(1 row)  
  
  
postgres=# delete from t_src where id=1;  
DELETE 0  
postgres=# select * from t_src;  
 id |               info               | is_del |          mod_time            
----+----------------------------------+--------+----------------------------  
  1 | 5bca407559081d6cfc1154fd0f17b6a9 | t      | 2018-05-12 08:54:43.158809  
(1 row)

7、创建压测脚本

vi test.sql  
  
\set id1 random(1,10000000)  
\set id2 random(1,20000000)  
insert into t_src values (:id1, md5(random()::text)) on conflict (id) do update set info=excluded.info;  
delete from t_src where id=:id2;

8、压测，高压插入、更新、删除动作，每秒处理13.8万行。

pgbench -M prepared -n -r -P 1 -f ./test.sql -c 64 -j 64 -T 120  
  
  
transaction type: ./test.sql  
scaling factor: 1  
query mode: prepared  
number of clients: 64  
number of threads: 64  
duration: 120 s  
number of transactions actually processed: 16634225  
latency average = 0.462 ms  
latency stddev = 0.506 ms  
tps = 138437.925513 (including connections establishing)  
tps = 138445.093708 (excluding connections establishing)  
statement latencies in milliseconds:  
         0.002  \set id1 random(1,10000000)  
         0.000  \set id2 random(1,20000000)  
         0.298  insert into t_src values (:id1, md5(random()::text)) on conflict (id) do update set info=excluded.info;  
         0.163  delete from t_src where id=:id2;

目标端

1、创建目标表

create table t_dst (  
  id int primary key,   
  info text not null,   
  is_del boolean default null,   
  mod_time timestamp not null default clock_timestamp()  
);

2、创建索引

create index idx_t_dst_1 on t_dst(mod_time);

数据同步

1、源端数据增量同步到目标端

(在同一DB中模拟，后面有例子讲源和目标在不同集群的DEMO)

do language plpgsql $$  
declare  
  pos timestamp;           -- 位点  
  pre interval := '10 s';  -- 缓冲10秒，防止空洞，根据业务层设置  
begin  
  -- 已同步位点  
  select max(mod_time) into pos from t_dst;  
    
  -- NULL表示目标端没有数据  
  if pos is null then  
    -- 缓冲上限   
    pos := now()::timestamp;  
    insert into t_dst select * from t_src where mod_time < (pos - pre) on conflict (id) do update set   -- on conflict 合并insert,update  
      info=excluded.info, is_del=excluded.is_del, mod_time=excluded.mod_time  
      where t_dst.info is distinct from excluded.info or  
      t_dst.is_del is distinct from excluded.is_del or  
      t_dst.mod_time is distinct from excluded.mod_time;  
    return;  
  end if;  
  
  -- 同步超过位点的数据  
  insert into t_dst select * from t_src where mod_time > pos and mod_time < (now()::timestamp - pre) on conflict (id) do update set   
    info=excluded.info, is_del=excluded.is_del, mod_time=excluded.mod_time  
    where t_dst.info is distinct from excluded.info or  
    t_dst.is_del is distinct from excluded.is_del or  
    t_dst.mod_time is distinct from excluded.mod_time;  
  return;  
end;  
$$;

2、一边压测，一边调用以上过程同步，最后达到一致性状态，检查一致性的SQL如下：

select sum(hashtext((t.*)::text)),count(*),sum(case is_del when true then 0 else 1 end) from t_src t;  
  
      sum       |  count  |   sum     
----------------+---------+---------  
 -2967631712018 | 8299587 | 6199359  
(1 row)  
  
  
select sum(hashtext((t.*)::text)),count(*),sum(case is_del when true then 0 else 1 end) from t_dst t;  
  
      sum       |  count  |   sum     
----------------+---------+---------  
 -2967631712018 | 8299587 | 6199359  
(1 row)

二、源PostgreSQL, 目标Greenplum, 增量复制delete,insert,update

第一种方法，DELETE使用逻辑标记，所以实际上数据并没有删除。

还有一种方法，可以把DELETE的记录，MOVE到另一张表。

方法与一差不多，只是把rule改一下，改成INSERT到其他表。

略

三、源PostgreSQL, 目标Greenplum, 增量复制insert,update

上游没有del，或者说不需要捕获DEL操作 (DEL操作，人为在上下游同时执行SQL来删除)

源端

1、创建表

create table tbl_src (  
  id int primary key,   
  info text not null,   
  mod_time timestamp not null default clock_timestamp()  -- update时间戳  
);

2、创建索引

create index idx_tbl_src_1 on tbl_src(mod_time);

3、创建触发器函数，更新时，自动更新时间戳字段

create or replace function tg1_tbl_src() returns trigger as $$  
declare  
begin  
  NEW.mod_time := clock_timestamp();  
  return NEW;  
end;  
$$ language plpgsql strict;

4、创建触发器

create trigger tg1 before update on tbl_src for each row execute procedure tg1_tbl_src();

5、压测

vi test.sql  
\set id1 random(1,10000000)  
insert into tbl_src values (:id1, md5(random()::text)) on conflict (id) do update set info=excluded.info;  
  
  
pgbench -M prepared -n -r -P 1 -f ./test.sql -c 64 -j 64 -T 120

目标端

1、创建目标表

create table tbl_dst (  
  id int primary key,   
  info text not null,   
  mod_time timestamp not null default clock_timestamp()  
);

2、创建索引

create index idx_tbl_dst_1 on tbl_dst(mod_time);

数据同步

在同一个实例中模拟。

1、调用过程，同步数据。

do language plpgsql $$  
declare  
  pos timestamp;  -- 位点  
  pre interval := '10 s';  -- 缓冲10秒，防止空洞，根据业务层设置  
begin  
  select max(mod_time) into pos from tbl_dst;  
  if pos is null then  
    pos := now()::timestamp;  
    insert into tbl_dst select * from tbl_src where mod_time < (pos - pre) on conflict (id) do update set   
      info=excluded.info, mod_time=excluded.mod_time  
      where tbl_dst.info is distinct from excluded.info or  
      tbl_dst.mod_time is distinct from excluded.mod_time;  
    return;  
  end if;  
  
  insert into tbl_dst select * from tbl_src where mod_time > pos and mod_time < (now()::timestamp - pre) on conflict (id) do update set   
    info=excluded.info, mod_time=excluded.mod_time  
    where tbl_dst.info is distinct from excluded.info or  
    tbl_dst.mod_time is distinct from excluded.mod_time;  
  return;  
end;  
$$;

2、一边压测，一边调用以上过程同步数据，最后检查一致性

select sum(hashtext((t.*)::text)),count(*) from tbl_src t;  
  
      sum      |  count    
---------------+---------  
 2739329060132 | 6725930  
(1 row)  
  
select sum(hashtext((t.*)::text)),count(*) from tbl_dst t;  
  
      sum      |  count    
---------------+---------  
 2739329060132 | 6725930  
(1 row)

四、源PostgreSQL, 目标Greenplum, 增量复制insert,update

同样，不包括DELETE的复制。只复制insert和update。

由于greenplum 没有 insert into on conflict 的功能，所以需要采用临时表，分步实现MERGE。

同步方法

1. 目标端，HDB PG, 获取max(mod_time)

2. 源端，PG, 拉取增量

3. 目标端，增量数据，导入HDB PG 临时表

4. 目标端，HDB PG ，DELETE from 目标表 using 临时表

5. 目标端，HDB PG ，insert into 目标表 select * from 临时表

6. 目标端，清空临时表

目标端

1、创建临时表

create table tbl_dst_tmp (  
  id int primary key,   
  info text not null,   
  mod_time timestamp not null default clock_timestamp()  
);

2、创建数据同步的脚本

vi imp.sh  
  
#!/bin/bash  
  
# 1. HDB PG, 获取max(mod_time)   
MOD_TIME=`PGPASSWORD="pwd1234" psql -h 127.0.0.1 -p 4000 -U postgres postgres -q -t -A -c "select coalesce(max(mod_time),'4714-11-24 00:00:00 BC'::timestamp) from tbl_dst"`  
  
# 2. PG, 拉取增量  
# 3. 增量数据，导入HDB PG 临时表  
# 使用 linux 管道同步上下游  
PGPASSWORD="pwd" psql -h 10.31.124.69 -p 4000 -U postgres postgres -c "copy (select * from tbl_src where mod_time > '$MOD_TIME'::timestamp and mod_time < (now()-'10 sec'::interval) ) to stdout" | PGPASSWORD="pwd1234" psql -h 127.0.0.1 -p 4000 -U postgres postgres -c "copy tbl_dst_tmp from stdin"  
  
# 4. HDB PG ，DELETE from 目标表 using 临时表   
# 5. HDB PG ，insert into 目标表 select * from 临时表   
# 6. 清空临时表   
# 放在一个事务中，并且对临时表加锁保护。  
PGPASSWORD="pwd1234" psql -h 127.0.0.1 -p 4000 -U postgres postgres <<EOF  
begin;  
lock table tbl_dst_tmp in ACCESS EXCLUSIVE mode;  
delete from tbl_dst using tbl_dst_tmp where tbl_dst.id=tbl_dst_tmp.id;  
insert into tbl_dst select * from tbl_dst_tmp;  
truncate tbl_dst_tmp;  
end;  
EOF  
  
  
chmod 700 imp.sh

3、一边压测，一边同步

COPY 6725930  
BEGIN  
LOCK TABLE  
DELETE 0  
INSERT 0 6725930  
TRUNCATE TABLE  
COMMIT  
  
  
COPY 0  
BEGIN  
LOCK TABLE  
DELETE 0  
INSERT 0 0  
TRUNCATE TABLE  
COMMIT  
  
  
COPY 78423  
BEGIN  
LOCK TABLE  
DELETE 52796  
INSERT 0 78423  
TRUNCATE TABLE  
COMMIT  
  
  
COPY 486817  
BEGIN  
LOCK TABLE  
DELETE 327603  
INSERT 0 486817  
TRUNCATE TABLE  
COMMIT  
  
  
COPY 2019059  
BEGIN  
LOCK TABLE  
DELETE 1381561  
INSERT 0 2019059  
TRUNCATE TABLE  
COMMIT

COPY 864687  
Timing is on.  
BEGIN  
Time: 0.149 ms  
LOCK TABLE  
Time: 0.629 ms  
DELETE 652117  
Time: 11029.147 ms (00:11.029)  
INSERT 0 864687  
Time: 10180.576 ms (00:10.181)  
TRUNCATE TABLE  
Time: 4.729 ms  
COMMIT  
Time: 53.555 ms  
  
  
  
COPY 2235200  
Timing is on.  
BEGIN  
Time: 0.178 ms  
LOCK TABLE  
Time: 0.702 ms  
DELETE 1719572  
Time: 18621.210 ms (00:18.621)  
INSERT 0 2235200  
Time: 27716.155 ms (00:27.716)  
TRUNCATE TABLE  
Time: 63.408 ms  
COMMIT  
Time: 81.915 ms  



COPY 5448790
Timing is on.
BEGIN
Time: 0.141 ms
LOCK TABLE
Time: 0.552 ms
DELETE 4486067
Time: 15297.884 ms (00:15.298)
INSERT 0 5448790
Time: 35654.723 ms (00:35.655)
TRUNCATE TABLE
Time: 1.860 ms
COMMIT
Time: 125.503 ms

4、检查数据一致性

postgres=# select sum(hashtext((t.*)::text)),count(*) from tbl_dst t;  
      sum       |  count    
----------------+---------  
 -1351270286348 | 7548269  
(1 row)  
  
postgres=# select sum(hashtext((t.*)::text)),count(*) from tbl_src t;  
      sum       |  count    
----------------+---------  
 -1351270286348 | 7548269  
(1 row)

五、源SQL Server, 目标Greenplum, 增量复制insert,update

SQL Server有一个时间戳类型timestamp，会自动记录数据插入，更新的时间，所以不需要用触发器来实现标记。

SQL Server的timestamp，每个表只能建一个，并且在表上绝对唯一。8个字节，是一个相对时间，可以与bigint互相转换。

目标端可以使用bigint来存储SQL Server的timestamp类型。

其他的方法与章节四类似，注意标记位的读取、比较时需要转换一下(timestamp, bigint)。 BIGINT 最小值为 (-9223372036854775808)::int8 。

DEMO1，位点从HDB PG读取。

1、HDB PG, 创建临时表

create table tbl_dst (  
  id int primary key,   
  info text not null,   
  mod_time int8 not null   -- 使用int8代替时间戳，对应ms sql的timestamp
); 

-- 临时表

create table tbl_dst_tmp (  
  id int primary key,   
  info text not null,   
  mod_time int8 not null   -- 使用int8代替时间戳，对应ms sql的timestamp
);

2、数据同步的流程（MS SQL -> HDB PG，无法封装到LINUX SHELL中，可能需要ETL程序介入。流程大致如下）

# 1. HDB PG, 获取max(mod_time)   
MOD_TIME=`PGPASSWORD="pwd1234" psql -h 127.0.0.1 -p 4000 -U postgres postgres -q -t -A -c "select coalesce(max(mod_time),(-9223372036854775808)::int8) from tbl_dst"`  
  
# 2. MS SQL, 拉取增量  
select id, info, convert(bigint, mod_time) from tbl_src where convert(bigint,mod_time) > $MOD_TIME;

# 3. 增量数据，导入HDB PG 临时表  

insert into tbl_dst_tmp ...从步骤2来的数据...;

## 可能的话，还是用COPY更快，"copy tbl_dst_tmp from stdin"  
  
# 4. HDB PG ，DELETE from 目标表 using 临时表   
# 5. HDB PG ，insert into 目标表 select * from 临时表   
# 6. 清空临时表   

# 放在一个事务中，并且对临时表加锁保护。  
PGPASSWORD="pwd1234" psql -h 127.0.0.1 -p 4000 -U postgres postgres <<EOF  
begin;  
lock table tbl_dst_tmp in ACCESS EXCLUSIVE mode;  
delete from tbl_dst using tbl_dst_tmp where tbl_dst.id=tbl_dst_tmp.id;  
insert into tbl_dst select * from tbl_dst_tmp;  
truncate tbl_dst_tmp;  
end;  
EOF

DEMO2, 位点记录在MS SQL里面，而不是从HDB PG读取。

MS SQL, 创建任务表

create table tbl_job (
  tbl_name name primary key,
  job_time datetime,
  offset_val_low bigint,
  offset_val_up bigint
)

从任务表选择上一次的最低位点

select offset_val_up into v1 from tbl_job where tbl_name='tbl_src'; 

如果上一次没有同步，则说明是第一次同步

if v1 is null then v1 := -1; end if; 

选择本次同步的最大位点

select max(convert(bigint, mod_time)) into v2 from tbl_src; 
  
更新任务表

insert or update into tbl_job values ('tbl_src', now(), v1, v2 ); 

导出本次的增量数据
  
select * from tbl_src where convert(bigint, mod_time) > v1 and convert(bigint, mod_time) <= v2;

或者你还可以试试其他ETL软件：

http://www.symmetricds.org/about/overview

https://github.com/pivotalguru/outsourcer 专业的sql server,oracle同步到greenplum的软件

商业软件：

https://dbconvert.com/mssql/postgresql/

https://www.convert-in.com/mss2pgs.htm#

xDB

https://www.enterprisedb.com/docs/en/6.2/repguide/EDB_Postgres_Replication_Server_Users_Guide.1.10.html#pID0E0HSK0HA