标签
PostgreSQL , 时序数据库 , 时序 , 滑动窗口 , 递归查询 , subquery , 窗口查询 , 求最新值
背景
在很多场景中，都会有数据合并、清洗的需求。
例如：
1、记录了表的变更明细(insert,update,delete)，需要合并明细，从明细中快速取到每个PK的最新值。
2、有很多传感器，不断的在上报数据，要快速的取出每个传感器的最新状态。
对于这类需求，可以使用窗口查询，但是如何加速，如何快速的取出批量数据呢？
PostgreSQL是*的开源数据库，优化方法之多，超乎你的想象。
时序数据取值优化
1、唯一值较少时，并且唯一值范围未知时，使用递归。
方法如下：
《PostgrSQL 递归SQL的几个应用 - 极客与正常人的思维》
《时序数据合并场景加速分析和实现 - 复合索引，窗口分组查询加速，变态递归加速》
2、唯一值较少时，并且唯一值范围确定时，使用subquery。
方法如下：
《PostgreSQL 海量时序数据(任意滑动窗口实时统计分析) - 传感器、人群、物体等对象跟踪》
3、唯一值较多时，1 使用窗口查询，比前面的方法更加适合。
4、唯一值较多时，2 使用流式计算，比方法3更加优秀。
方法如下：
《(流式、lambda、触发器)实时处理大比拼 - 物联网(IoT)\金融,时序处理最佳实践》
本文将对前三种方法做一个比较。
方法4 流计算就不用比了，因为它什么时候都是最强大的。通杀所有场景。等pipelineDB 插件化吧，阿里云RDS PG 10会集成pipelineDB的功能。
递归 vs subquery vs 窗口
以500万数据为例，对比这几种方法的适应场景。
1 递归
场景一、唯一值较多(100万唯一值)
1、建表
\timing
drop table test;
create unlogged table test(id int , info text, crt_time timestamp);
2、构造数据
insert into test select ceil(random()*1000000), md5(random()::text), clock_timestamp() from generate_series(1,5000000);
3、创建索引
create index idx_test_1 on test (id, crt_time desc);
4、递归查询效率
explain (analyze,verbose,timing,costs,buffers) with recursive skip as (
(
select test as v from test where id in (select id from test where id is not null order by id,crt_time desc limit 1) limit 1
)
union all
(
select (
select t as v from test t where t.id>(s.v).id and t.id is not null order by id,crt_time desc limit 1
) from skip s where (s.v).id is not null
) -- 这里的where (s.v).id is not null 一定要加, 否则就死循环了.
)
select (t.v).id, (t.v).info, (t.v).crt_time from skip t where t.* is not null;
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
CTE Scan on skip t (cost=54.35..56.37 rows=100 width=44) (actual time=0.042..6626.084 rows=993288 loops=1)
Output: (t.v).id, (t.v).info, (t.v).crt_time
Filter: (t.* IS NOT NULL)
Rows Removed by Filter: 1
Buffers: shared hit=3976934
CTE skip
-> Recursive Union (cost=0.91..54.35 rows=101 width=69) (actual time=0.034..6006.615 rows=993289 loops=1)
Buffers: shared hit=3976934
-> Limit (cost=0.91..0.93 rows=1 width=69) (actual time=0.033..0.033 rows=1 loops=1)
Output: test.*
Buffers: shared hit=8
-> Nested Loop (cost=0.91..10.19 rows=500 width=69) (actual time=0.032..0.032 rows=1 loops=1)
Output: test.*
Buffers: shared hit=8
-> HashAggregate (cost=0.48..0.49 rows=1 width=4) (actual time=0.021..0.021 rows=1 loops=1)
Output: test_1.id
Group Key: test_1.id
Buffers: shared hit=4
-> Limit (cost=0.43..0.47 rows=1 width=12) (actual time=0.016..0.016 rows=1 loops=1)
Output: test_1.id, test_1.crt_time
Buffers: shared hit=4
-> Index Only Scan using idx_test_1 on public.test test_1 (cost=0.43..173279.36 rows=5000002 width=12) (actual time=0.015..0.015 rows=1 loops=1)
Output: test_1.id, test_1.crt_time
Index Cond: (test_1.id IS NOT NULL)
Heap Fetches: 1
Buffers: shared hit=4
-> Index Scan using idx_test_1 on public.test (cost=0.43..9.64 rows=6 width=73) (actual time=0.009..0.009 rows=1 loops=1)
Output: test.*, test.id
Index Cond: (test.id = test_1.id)
Buffers: shared hit=4
-> WorkTable Scan on skip s (cost=0.00..5.14 rows=10 width=32) (actual time=0.006..0.006 rows=1 loops=993289)
Output: (SubPlan 1)
Filter: ((s.v).id IS NOT NULL)
Rows Removed by Filter: 0
Buffers: shared hit=3976926
SubPlan 1
-> Limit (cost=0.43..0.49 rows=1 width=81) (actual time=0.005..0.005 rows=1 loops=993288)
Output: t_1.*, t_1.id, t_1.crt_time
Buffers: shared hit=3976926
-> Index Scan using idx_test_1 on public.test t_1 (cost=0.43..102425.17 rows=1666667 width=81) (actual time=0.005..0.005 rows=1 loops=993288)
Output: t_1.*, t_1.id, t_1.crt_time
Index Cond: ((t_1.id > (s.v).id) AND (t_1.id IS NOT NULL))
Buffers: shared hit=3976926
Planning time: 0.354 ms
Execution time: 6706.105 ms
(45 rows)
场景二、唯一值较少(1000唯一值)
1、建表
\timing
drop table test;
create unlogged table test(id int , info text, crt_time timestamp);
2、构造数据
insert into test select ceil(random()*1000), md5(random()::text), clock_timestamp() from generate_series(1,5000000);
3、创建索引
create index idx_test_1 on test (id, crt_time desc);
4、递归查询效率
查询语句不变
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
CTE Scan on skip t (cost=55.09..57.11 rows=100 width=44) (actual time=0.046..8.859 rows=1000 loops=1)
Output: (t.v).id, (t.v).info, (t.v).crt_time
Filter: (t.* IS NOT NULL)
Rows Removed by Filter: 1
Buffers: shared hit=4007
CTE skip
-> Recursive Union (cost=0.91..55.09 rows=101 width=69) (actual time=0.039..8.203 rows=1001 loops=1)
Buffers: shared hit=4007
-> Limit (cost=0.91..1.67 rows=1 width=69) (actual time=0.038..0.038 rows=1 loops=1)
Output: test.*
Buffers: shared hit=8
-> Nested Loop (cost=0.91..6335.47 rows=8333 width=69) (actual time=0.038..0.038 rows=1 loops=1)
Output: test.*
Buffers: shared hit=8
-> HashAggregate (cost=0.48..0.49 rows=1 width=4) (actual time=0.021..0.021 rows=1 loops=1)
Output: test_1.id
Group Key: test_1.id
Buffers: shared hit=4
-> Limit (cost=0.43..0.47 rows=1 width=12) (actual time=0.016..0.017 rows=1 loops=1)
Output: test_1.id, test_1.crt_time
Buffers: shared hit=4
-> Index Only Scan using idx_test_1 on public.test test_1 (cost=0.43..173279.55 rows=5000002 width=12) (actual time=0.015..0.015 rows=1 loops=1)
Output: test_1.id, test_1.crt_time
Index Cond: (test_1.id IS NOT NULL)
Heap Fetches: 1
Buffers: shared hit=4
-> Index Scan using idx_test_1 on public.test (cost=0.43..6284.98 rows=5000 width=73) (actual time=0.015..0.015 rows=1 loops=1)
Output: test.*, test.id
Index Cond: (test.id = test_1.id)
Buffers: shared hit=4
-> WorkTable Scan on skip s (cost=0.00..5.14 rows=10 width=32) (actual time=0.008..0.008 rows=1 loops=1001)
Output: (SubPlan 1)
Filter: ((s.v).id IS NOT NULL)
Rows Removed by Filter: 0
Buffers: shared hit=3999
SubPlan 1
-> Limit (cost=0.43..0.49 rows=1 width=81) (actual time=0.007..0.007 rows=1 loops=1000)
Output: t_1.*, t_1.id, t_1.crt_time
Buffers: shared hit=3999
-> Index Scan using idx_test_1 on public.test t_1 (cost=0.43..102425.80 rows=1666667 width=81) (actual time=0.007..0.007 rows=1 loops=1000)
Output: t_1.*, t_1.id, t_1.crt_time
Index Cond: ((t_1.id > (s.v).id) AND (t_1.id IS NOT NULL))
Buffers: shared hit=3999
Planning time: 0.353 ms
Execution time: 8.980 ms
(45 rows)
2 subquery
场景一、唯一值较多(100万唯一值)
1、subquery查询效率
如果ID的取值范围特别广，SUBQUERY就很不划算。
需要维护一张唯一ID表，这里使用generate_series来代替这张表，用于测试。
explain (analyze,verbose,timing,costs,buffers) select (select test from test where id=t.id order by crt_time desc limit 1) from generate_series(1,1000000) t(id);
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------
Function Scan on pg_catalog.generate_series t (cost=0.00..1976.65 rows=1000 width=32) (actual time=70.682..2835.109 rows=1000000 loops=1)
Output: (SubPlan 1)
Function Call: generate_series(1, 1000000)
Buffers: shared hit=3997082
SubPlan 1
-> Limit (cost=0.43..1.97 rows=1 width=77) (actual time=0.002..0.002 rows=1 loops=1000000)
Output: test.*, test.crt_time
Buffers: shared hit=3997082
-> Index Scan using idx_test_1 on public.test (cost=0.43..9.64 rows=6 width=77) (actual time=0.002..0.002 rows=1 loops=1000000)
Output: test.*, test.crt_time
Index Cond: (test.id = t.id)
Buffers: shared hit=3997082
Planning time: 0.119 ms
Execution time: 2892.712 ms
(14 rows)
场景二、唯一值较少(1000唯一值)
1、subquery查询效率
查询语句有变
explain (analyze,verbose,timing,costs,buffers) select (select test from test where id=t.id order by crt_time desc limit 1) from generate_series(1,1000) t(id);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Function Scan on pg_catalog.generate_series t (cost=0.00..1699.41 rows=1000 width=32) (actual time=0.107..7.041 rows=1000 loops=1)
Output: (SubPlan 1)
Function Call: generate_series(1, 1000)
Buffers: shared hit=4000
SubPlan 1
-> Limit (cost=0.43..1.69 rows=1 width=77) (actual time=0.006..0.007 rows=1 loops=1000)
Output: test.*, test.crt_time
Buffers: shared hit=4000
-> Index Scan using idx_test_1 on public.test (cost=0.43..6284.98 rows=5000 width=77) (actual time=0.006..0.006 rows=1 loops=1000)
Output: test.*, test.crt_time
Index Cond: (test.id = t.id)
Buffers: shared hit=4000
Planning time: 0.131 ms
Execution time: 7.126 ms
(14 rows)
3 窗口查询
场景一、唯一值较多(100万唯一值)
1、窗口查询效率
explain (analyze,verbose,timing,costs,buffers) select id,info,crt_time from (select row_number() over (partition by id order by crt_time desc) as rn, * from test) t where rn=1;
postgres=# explain (analyze,verbose,timing,costs,buffers) select id,info,crt_time from (select row_number() over (partition by id order by crt_time desc) as rn, * from test) t where rn=1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------
Subquery Scan on t (cost=0.43..310779.41 rows=25000 width=45) (actual time=0.027..6398.308 rows=993288 loops=1)
Output: t.id, t.info, t.crt_time
Filter: (t.rn = 1)
Rows Removed by Filter: 4006712
Buffers: shared hit=5018864
-> WindowAgg (cost=0.43..248279.39 rows=5000002 width=53) (actual time=0.026..5973.497 rows=5000000 loops=1)
Output: row_number() OVER (?), test.id, test.info, test.crt_time
Buffers: shared hit=5018864
-> Index Scan using idx_test_1 on public.test (cost=0.43..160779.35 rows=5000002 width=45) (actual time=0.019..4058.476 rows=5000000 loops=1)
Output: test.id, test.info, test.crt_time
Buffers: shared hit=5018864
Planning time: 0.121 ms
Execution time: 6446.901 ms
(13 rows)
场景二、唯一值较少(1000唯一值)
1、窗口查询效率
查询语句不变
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------
Subquery Scan on t (cost=0.43..310779.61 rows=25000 width=45) (actual time=0.027..6176.801 rows=1000 loops=1)
Output: t.id, t.info, t.crt_time
Filter: (t.rn = 1)
Rows Removed by Filter: 4999000
Buffers: shared hit=4744850 read=18157
-> WindowAgg (cost=0.43..248279.58 rows=5000002 width=53) (actual time=0.026..5822.576 rows=5000000 loops=1)
Output: row_number() OVER (?), test.id, test.info, test.crt_time
Buffers: shared hit=4744850 read=18157
-> Index Scan using idx_test_1 on public.test (cost=0.43..160779.55 rows=5000002 width=45) (actual time=0.020..4175.082 rows=5000000 loops=1)
Output: test.id, test.info, test.crt_time
Buffers: shared hit=4744850 read=18157
Planning time: 0.108 ms
Execution time: 6176.924 ms
(13 rows)
横向效率对比图
| 数据量 | 唯一值个数 | 窗口查询(ms) | subquery(ms) | 递归查询(ms) |
|---|---|---|---|---|
| 500 万 | 100万 | 6446 | 2892 | 6706 |
| 500 万 | 100万 | 6176 | 7 | 9 |
云端产品
类似案例
《PostgrSQL 递归SQL的几个应用 - 极客与正常人的思维》
《时序数据合并场景加速分析和实现 - 复合索引，窗口分组查询加速，变态递归加速》
《(流式、lambda、触发器)实时处理大比拼 - 物联网(IoT)\金融,时序处理最佳实践》
《PostgreSQL 海量时序数据(任意滑动窗口实时统计分析) - 传感器、人群、物体等对象跟踪》
《车联网案例，轨迹清洗 - 阿里云RDS PostgreSQL最佳实践 - 窗口函数》
《PostgreSQL 时序最佳实践 - 证券交易系统数据库设计 - 阿里云RDS PostgreSQL最佳实践》
小结
随着物联网的发展，越来越多的时序数据产生，求时序数据的最新值，滑动窗口内的最新值，已经成为时序业务里面非常场景的需求。
PostgreSQL是最先进的开源数据库，它不像很多数据库解决问题往往只有一种方法。PostgreSQL解决一个问题有多种方法，什么方法最优，就看你对它的了解了。
1、唯一值较少时，并且唯一值范围未知时，使用递归。
2、唯一值较少时，并且唯一值范围确定时(例如范围是100万，但是此批数据只出现了50万个，那么如果你有这50万个的ID，性能是最好的，否则需要扫100万。例如用户数是1亿，一个区间的活跃用户可能只有几万。)，使用subquery。
3、唯一值较多时，1 使用窗口查询，比前面的方法更加适合。
4、唯一值较多时，2 使用流式计算，比方法3更加优秀。