Greenplum merge insert 用法与性能 (insert on conflict) - 1
背景
PostgreSQL insert on conflict语法非常强大，支持合并写入（当违反某唯一约束时，冲突则更新，不冲突则写入），同时支持流式计算。
流计算例子链接：
《PostgreSQL 流式统计 - insert on conflict 实现 流式 UV(distinct), min, max, avg, sum, count …》
| [《HTAP数据库 PostgreSQL 场景与性能测试之 22 - (OLTP) merge insert | upsert | insert on conflict | 合并写入》](../201711/20171107_23.md) |
《PostgreSQL upsert功能(insert on conflict do)的用法》
《PostgreSQL 10.0 preview 功能增强 - 支持分区表ON CONFLICT .. DO NOTHING》
PostgreSQL insert on conflict语法如下：
Command: INSERT
Description: create new rows in a table
Syntax:
[ WITH [ RECURSIVE ] with_query [, ...] ]
INSERT INTO table_name [ AS alias ] [ ( column_name [, ...] ) ]
[ OVERRIDING { SYSTEM | USER} VALUE ]
{ DEFAULT VALUES | VALUES ( { expression | DEFAULT } [, ...] ) [, ...] | query }
[ ON CONFLICT [ conflict_target ] conflict_action ]
[ RETURNING * | output_expression [ [ AS ] output_name ] [, ...] ]
where conflict_target can be one of:
( { index_column_name | ( index_expression ) } [ COLLATE collation ] [ opclass ] [, ...] ) [ WHERE index_predicate ]
ON CONSTRAINT constraint_name
and conflict_action is one of:
DO NOTHING
DO UPDATE SET { column_name = { expression | DEFAULT } |
( column_name [, ...] ) = [ ROW ] ( { expression | DEFAULT } [, ...] ) |
( column_name [, ...] ) = ( sub-SELECT )
} [, ...]
[ WHERE condition ]
Greenplum的版本较低，还不支持insert on conflict的语法。
如果需要在Greenplum中实现类似的功能该如何操作？
DEMO
ID为PK，以它为合并列，举例。
1、目标表，也就是需要合并写入的目标：
create table t( id int primary key, c1 int , c2 int, c3 int, c4 int, c5 int, crt_time timestamp);
2、中间表，也就是用户只管插入的表：
create table t_tmp(like t);
写入一些中间记录。
insert into t_tmp values(1,1,2,3,null,null,now());
insert into t_tmp values(1,1,2,4,null,null,now());
insert into t_tmp values(1,1,2,3,null,7,now());
insert into t_tmp values(1,1,null,3,5,6,now());
postgres=# select * from t_tmp;
id | c1 | c2 | c3 | c4 | c5 | crt_time
----+----+----+----+----+----+----------------------------
1 | 1 | 2 | 3 | | | 2017-12-13 17:03:16.28482
1 | 1 | 2 | 4 | | | 2017-12-13 17:03:16.286302
1 | 1 | 2 | 3 | | 7 | 2017-12-13 17:03:16.635121
1 | 1 | | 3 | 5 | 6 | 2017-12-13 17:03:25.434191
(4 rows)
3、窗口合并，按唯一值约束，仅提取一条（可能存在窗口内合并的需求，例如按时间取最新，比如以最后一条为准，又或者以有值，且最新的为准）。
以有值切最新为准例子：
select distinct on (id)
id,
first_value(c1) over (partition by id order by (case when c1 is null then null else crt_time end) desc nulls last) as c1,
first_value(c2) over (partition by id order by (case when c2 is null then null else crt_time end) desc nulls last) as c2,
first_value(c3) over (partition by id order by (case when c3 is null then null else crt_time end) desc nulls last) as c3,
first_value(c4) over (partition by id order by (case when c4 is null then null else crt_time end) desc nulls last) as c4,
first_value(c5) over (partition by id order by (case when c5 is null then null else crt_time end) desc nulls last) as c5,
first_value(crt_time) over (partition by id order by crt_time desc) as crt_time
from t_tmp ;
id | c1 | c2 | c3 | c4 | c5 | crt_time
----+----+----+----+----+----+----------------------------
1 | 1 | 2 | 3 | 5 | 6 | 2017-12-13 17:03:25.434191
(1 row)
存储中间结果：
create table t_tmp1 (like t) ;
insert into t_tmp1
select distinct on (id)
id,
first_value(c1) over (partition by id order by (case when c1 is null then null else crt_time end) desc nulls last) as c1,
first_value(c2) over (partition by id order by (case when c2 is null then null else crt_time end) desc nulls last) as c2,
first_value(c3) over (partition by id order by (case when c3 is null then null else crt_time end) desc nulls last) as c3,
first_value(c4) over (partition by id order by (case when c4 is null then null else crt_time end) desc nulls last) as c4,
first_value(c5) over (partition by id order by (case when c5 is null then null else crt_time end) desc nulls last) as c5,
first_value(crt_time) over (partition by id order by crt_time desc) as crt_time
from t_tmp ;
4、合并写入：
将窗口提取的结果，合并写入目标表。
4.1、INNER JOIN，覆盖旧记录，同时补齐旧的字段（以NULL为判断条件。如果新的记录没有值，则取旧记录的值。）提取。
create table t_tmp2 (like t);
insert into t_tmp2
select
t_tmp.id,
coalesce(t_tmp.c1, t.c1),
coalesce(t_tmp.c2, t.c2),
coalesce(t_tmp.c3, t.c3),
coalesce(t_tmp.c4, t.c4),
coalesce(t_tmp.c5, t.c5),
coalesce(t_tmp.crt_time, t.crt_time)
from
t_tmp1 as t_tmp
inner join
t
using (id);
4.2、DELETE USING，删除全量表的符合条件的记录。
delete from t using t_tmp2 where t.id=t_tmp2.id;
4.3、INSERT
insert into t
select t_tmp1.* from t_tmp1 left join t_tmp2 using (id) where t_tmp2.* is null
union all
select * from t_tmp2;
Greenplum merge insert 性能
硬件：使用一台64线程机器，单机启动48个segment。
1、全量数据20亿。
create table t(id int, c1 int , c2 int, c3 int, c4 int, c5 int, crt_time timestamp) with (APPENDONLY=true, ORIENTATION=column);
insert into t select id, null,null,null,null,10000, now() from generate_series(1,2000000000) t(id);
2、增量数据1000万条，涉及500万个ID。
create table t_tmp(like t);
insert into t_tmp select random()*1000000, random()*100,null,null,null,null, clock_timestamp() from generate_series(1,2000000) t(id);
insert into t_tmp select random()*2000000, null,random()*100,null,null,null, clock_timestamp() from generate_series(1,2000000) t(id);
insert into t_tmp select random()*3000000, null,null,random()*100,null,null, clock_timestamp() from generate_series(1,2000000) t(id);
insert into t_tmp select random()*4000000, null,null,null,random()*100,null, clock_timestamp() from generate_series(1,2000000) t(id);
insert into t_tmp select random()*5000000, null,null,null,null,random()*100, clock_timestamp() from generate_series(1,2000000) t(id);
总耗时4.5秒。
3、合并。
增量数据，窗口合并去重。
create table t_tmp1 (like t) ;
insert into t_tmp1
select distinct on (id)
id,
first_value(c1) over (partition by id order by ((case when c1 is null then null else crt_time end) is null), (case when c1 is null then null else crt_time end) desc) as c1,
first_value(c2) over (partition by id order by ((case when c2 is null then null else crt_time end) is null), (case when c2 is null then null else crt_time end) desc) as c2,
first_value(c3) over (partition by id order by ((case when c3 is null then null else crt_time end) is null), (case when c3 is null then null else crt_time end) desc) as c3,
first_value(c4) over (partition by id order by ((case when c4 is null then null else crt_time end) is null), (case when c4 is null then null else crt_time end) desc) as c4,
first_value(c5) over (partition by id order by ((case when c5 is null then null else crt_time end) is null), (case when c5 is null then null else crt_time end) desc) as c5,
first_value(crt_time) over (partition by id order by crt_time desc) as crt_time
from t_tmp ;
INSERT 0 3628283
Time: 5208.968 ms
使用增量数据，提取并合并旧数据。
create table t_tmp2 (like t);
insert into t_tmp2
select
t_tmp.id,
coalesce(t_tmp.c1, t.c1),
coalesce(t_tmp.c2, t.c2),
coalesce(t_tmp.c3, t.c3),
coalesce(t_tmp.c4, t.c4),
coalesce(t_tmp.c5, t.c5),
coalesce(t_tmp.crt_time, t.crt_time)
from
t_tmp1 as t_tmp
inner join
t
using (id);
INSERT 0 3628282
Time: 9504.092 ms
删除旧数据。
delete from t using t_tmp2 where t.id=t_tmp2.id;
DELETE 3628282
Time: 15356.920 ms
插入新增、以及合并的增量数据。
insert into t
select t_tmp1.* from t_tmp1 left join t_tmp2 using (id) where t_tmp2.* is null
union all
select * from t_tmp2;
INSERT 0 3628283
Time: 778.014 ms
数据校验
-- 中间结果
postgres=# select * from t_tmp where id=9;
id | c1 | c2 | c3 | c4 | c5 | crt_time
----+----+----+----+----+----+----------------------------
9 | 99 | | | | | 2017-12-13 23:18:14.65243
9 | 7 | | | | | 2017-12-13 23:18:14.817107
9 | | 9 | | | | 2017-12-13 23:18:15.292311
9 | | 56 | | | | 2017-12-13 23:18:15.449415
(4 rows)
-- 中间结果
postgres=# select * from t_tmp where id=446;
id | c1 | c2 | c3 | c4 | c5 | crt_time
-----+----+----+----+----+----+----------------------------
446 | 43 | | | | | 2017-12-13 23:18:14.291335
446 | 16 | | | | | 2017-12-13 23:18:14.715026
446 | 45 | | | | | 2017-12-13 23:18:15.048879
446 | | 34 | | | | 2017-12-13 23:18:15.646904
446 | | 7 | | | | 2017-12-13 23:18:15.81838
446 | | | 12 | | | 2017-12-13 23:18:16.220083
446 | | | 22 | | | 2017-12-13 23:18:16.26496
446 | | | | 97 | | 2017-12-13 23:18:17.464355
446 | | | | | 56 | 2017-12-13 23:18:18.427068
(9 rows)
-- 使用窗口合并后结果
postgres=# select * from t_tmp1 limit 10;
id | c1 | c2 | c3 | c4 | c5 | crt_time
-----+----+----+----+----+----+----------------------------
9 | 7 | 56 | | | | 2017-12-13 23:18:15.449415 -- 验证
25 | 69 | | 1 | | | 2017-12-13 23:18:16.161339
169 | 74 | 33 | 3 | | | 2017-12-13 23:18:16.71554
185 | 22 | | | | | 2017-12-13 23:18:14.93206
217 | 11 | 20 | 26 | | 59 | 2017-12-13 23:18:17.911174
270 | 55 | | 42 | | | 2017-12-13 23:18:16.494782
286 | 65 | 77 | 17 | | 75 | 2017-12-13 23:18:17.895121
430 | 12 | | 56 | | | 2017-12-13 23:18:16.744847
446 | 45 | 7 | 22 | 97 | 56 | 2017-12-13 23:18:18.427068 -- 验证
478 | 23 | 56 | | 25 | 77 | 2017-12-13 23:18:18.293153
(10 rows)
-- 合并到全量表后结果
postgres=# select * from t where id=9;
id | c1 | c2 | c3 | c4 | c5 | crt_time
----+----+----+----+----+-------+----------------------------
9 | 7 | 56 | | | 10000 | 2017-12-13 23:18:15.449415
(1 row)
postgres=# select * from t where id=446;
id | c1 | c2 | c3 | c4 | c5 | crt_time
-----+----+----+----+----+----+----------------------------
446 | 45 | 7 | 22 | 97 | 56 | 2017-12-13 23:18:18.427068
(1 row)
4、合并总耗时：
35秒
5、耗时分布
增量数据1000万条，涉及500万个ID。
4.5秒
增量数据，窗口合并去重。
5.2秒
使用增量数据，提取并合并旧数据。
9.5秒
删除旧数据。
15秒
插入新增、以及合并的增量数据。
0.7秒
Greenplum merge insert 限制
比较复杂、而且不支持新值使用NULL值(要支持的话，得修改一下覆盖逻辑)。
参考
《Greenplum 排序nulls first|last的 SQL写法实现》
《PostgreSQL 流式统计 - insert on conflict 实现 流式 UV(distinct), min, max, avg, sum, count …》
| [《HTAP数据库 PostgreSQL 场景与性能测试之 22 - (OLTP) merge insert | upsert | insert on conflict | 合并写入》](../201711/20171107_23.md) |
《PostgreSQL upsert功能(insert on conflict do)的用法》
《PostgreSQL 10.0 preview 功能增强 - 支持分区表ON CONFLICT .. DO NOTHING》
《PostgreSQL、Greenplum DML合并操作 最佳实践》