Digoal.zhou

Digoal.zhou

公益是一辈子的事, I'm digoal, Just do it.

Greenplum 行存、列存，堆表、AO表性能对比 - 阿里云HDB for PostgreSQL最佳实践

7 minute read

背景

《Greenplum 行存、列存，堆表、AO表的原理和选择》

以上文档详细的介绍了行存、列存，堆表、AO表的原理以及选择的依据。

《一个简单算法可以帮助物联网,金融 用户 节约98%的数据存储成本 (PostgreSQL,Greenplum帮你做到)》

以上文档介绍了提升基于列存的全局数据压缩比的方法。

《解密上帝之手 - 阿里云HDB for PostgreSQL数据库metascan特性(存储级、块级、batch级过滤与数据编排)》

以上文档介绍了局部编排，以及阿里云HDB for PostgreSQL数据库的metascan特性，（在不需要索引的情况下，如何提升任意列的选择性）。

压缩实际上是计算(CPU)换空间(磁盘)的做法，该不该做，我们还是先看看压缩比、性能损耗吧。

堆表

postgres=# create table t_heap(id int, c1 text, c2 int);  
CREATE TABLE  
  
postgres=# insert into t_heap select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
INSERT 0 10000000  
Time: 120526.098 ms

某个维度count查询。

postgres=# explain analyze select c2,count(*) from t_heap group by c2;  
                                                                          QUERY PLAN                                                                             
---------------------------------------------------------------------------------------------------------------------------------------------------------------  
 Gather Motion 48:1  (slice2; segments: 48)  (cost=1578949.03..1579074.98 rows=10076 width=12)  
   Rows out:  10001 rows at destination with 1354 ms to end, start offset by 519 ms.  
   ->  HashAggregate  (cost=1578949.03..1579074.98 rows=210 width=12)  
         Group By: t_heap.c2  
         Rows out:  Avg 208.4 rows x 48 workers.  Max 223 rows (seg17) with 0.001 ms to first row, 692 ms to end, start offset by 581 ms.  
         ->  Redistribute Motion 48:48  (slice1; segments: 48)  (cost=1578596.37..1578797.89 rows=210 width=12)  
               Hash Key: t_heap.c2  
               Rows out:  Avg 10001.0 rows x 48 workers at destination.  Max 10704 rows (seg17) with 596 ms to end, start offset by 581 ms.  
               ->  HashAggregate  (cost=1578596.37..1578596.37 rows=210 width=12)  
                     Group By: t_heap.c2  
                     Rows out:  Avg 10001.0 rows x 48 workers.  Max 10001 rows (seg0) with 0.006 ms to first row, 131 ms to end, start offset by 566 ms.  
                     ->  Seq Scan on t_heap  (cost=0.00..1528595.58 rows=208337 width=4)  
                           Rows out:  Avg 208333.3 rows x 48 workers.  Max 208401 rows (seg18) with 26 ms to first row, 901 ms to end, start offset by 573 ms.  
 Slice statistics:  
   (slice0)    Executor memory: 359K bytes.  
   (slice1)    Executor memory: 724K bytes avg x 48 workers, 724K bytes max (seg0).  
   (slice2)    Executor memory: 388K bytes avg x 48 workers, 388K bytes max (seg0).  
 Statement statistics:  
   Memory used: 128000K bytes  
 Settings:  optimizer=off  
 Optimizer status: legacy query optimizer  
 Total runtime: 1874.143 ms  
(22 rows)  
  
Time: 1879.480 ms

无索引，某个单值查询

postgres=# explain analyze select * from t_heap where c2=1;  
                                                             QUERY PLAN                                                               
------------------------------------------------------------------------------------------------------------------------------------  
 Gather Motion 48:1  (slice1; segments: 48)  (cost=0.00..1553595.98 rows=980 width=40)  
   Rows out:  1001 rows at destination with 1489 ms to end, start offset by 1.419 ms.  
   ->  Seq Scan on t_heap  (cost=0.00..1553595.98 rows=21 width=40)  
         Filter: c2 = 1  
         Rows out:  Avg 20.9 rows x 48 workers.  Max 35 rows (seg6) with 21 ms to first row, 613 ms to end, start offset by 853 ms.  
 Slice statistics:  
   (slice0)    Executor memory: 295K bytes.  
   (slice1)    Executor memory: 230K bytes avg x 48 workers, 230K bytes max (seg0).  
 Statement statistics:  
   Memory used: 128000K bytes  
 Settings:  optimizer=off  
 Optimizer status: legacy query optimizer  
 Total runtime: 1490.889 ms  
(13 rows)  
  
Time: 1492.516 ms

有索引（IO放大），某个单值查询

create index idx_t_heap on t_heap(c2);  
  
explain analyze select * from t_heap where c2=1;  
  
postgres=# explain analyze select * from t_heap where c2=1;  
                                                               QUERY PLAN                                                                 
----------------------------------------------------------------------------------------------------------------------------------------  
 Gather Motion 48:1  (slice1; segments: 48)  (cost=107.99..95579.14 rows=980 width=40)  
   Rows out:  1001 rows at destination with 34 ms to end, start offset by 1.331 ms.  
   ->  Bitmap Heap Scan on t_heap  (cost=107.99..95579.14 rows=21 width=40)  
         Recheck Cond: c2 = 1  
         Rows out:  Avg 20.9 rows x 48 workers.  Max 35 rows (seg6) with 0.236 ms to first row, 0.959 ms to end, start offset by 19 ms.  
         ->  Bitmap Index Scan on idx_t_heap  (cost=0.00..107.74 rows=21 width=0)  
               Index Cond: c2 = 1  
               Bitmaps out:  Avg 1.0 x 48 workers.  Max 1 (seg0) with 0.169 ms to end, start offset by 19 ms.  
               Work_mem used:  168K bytes avg, 282K bytes max (seg6).  
 Slice statistics:  
   (slice0)    Executor memory: 303K bytes.  
   (slice1)    Executor memory: 901K bytes avg x 48 workers, 901K bytes max (seg0).  Work_mem: 282K bytes max.  
 Statement statistics:  
   Memory used: 128000K bytes  
 Settings:  optimizer=off  
 Optimizer status: legacy query optimizer  
 Total runtime: 35.093 ms  
(17 rows)  
  
Time: 37.198 ms

APPENDONLY 行存储、列存储

测试脚本如下

vi test.sql  
  
\timing  
  
-- 堆表  
create table t_heap(id int, c1 text, c2 int);  
  
insert into t_heap select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_heap group by c2;  
  
explain analyze select * from t_heap where c2=1;  
  
create index idx_t_heap on t_heap(c2);  
  
explain analyze select * from t_heap where c2=1;  
  
-- AO 行存  
-- 不压缩, 8K  
  
create table t_ao_row_8k_0(id int, c1 text, c2 int) with (appendonly=true, blocksize=8192, ORIENTATION=row, COMPRESSTYPE=NONE, CHECKSUM=false);  
  
insert into t_ao_row_8k_0 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_row_8k_0 group by c2;  
  
explain analyze select * from t_ao_row_8k_0 where c2=1;  
  
create index idx_t_ao_row_8k_0 on t_ao_row_8k_0(c2);  
  
explain analyze select * from t_ao_row_8k_0 where c2=1;  
  
-- 不压缩, 2M  
  
create table t_ao_row_2m_0(id int, c1 text, c2 int) with (appendonly=true, blocksize=2097152, ORIENTATION=row, COMPRESSTYPE=NONE, CHECKSUM=false);  
  
insert into t_ao_row_2m_0 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_row_2m_0 group by c2;  
  
explain analyze select * from t_ao_row_2m_0 where c2=1;  
  
create index idx_t_ao_row_2m_0 on t_ao_row_2m_0(c2);  
  
explain analyze select * from t_ao_row_2m_0 where c2=1;  
  
-- 压缩比5, 8K  
  
create table t_ao_row_8k_5(id int, c1 text, c2 int) with (appendonly=true, blocksize=8192, ORIENTATION=row, COMPRESSTYPE=zlib, COMPRESSLEVEL=5, CHECKSUM=false);  
  
insert into t_ao_row_8k_5 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_row_8k_5 group by c2;  
  
explain analyze select * from t_ao_row_8k_5 where c2=1;  
  
create index idx_t_ao_row_8k_5 on t_ao_row_8k_5(c2);  
  
explain analyze select * from t_ao_row_8k_5 where c2=1;  
  
-- 压缩比5, 2M  
  
create table t_ao_row_2m_5(id int, c1 text, c2 int) with (appendonly=true, blocksize=2097152, ORIENTATION=row, COMPRESSTYPE=zlib, COMPRESSLEVEL=5, CHECKSUM=false);  
  
insert into t_ao_row_2m_5 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_row_2m_5 group by c2;  
  
explain analyze select * from t_ao_row_2m_5 where c2=1;  
  
create index idx_t_ao_row_2m_5 on t_ao_row_2m_5(c2);  
  
explain analyze select * from t_ao_row_2m_5 where c2=1;  
  
-- 压缩比9, 8K  
  
create table t_ao_row_8k_9(id int, c1 text, c2 int) with (appendonly=true, blocksize=8192, ORIENTATION=row, COMPRESSTYPE=zlib, COMPRESSLEVEL=9, CHECKSUM=false);  
  
insert into t_ao_row_8k_9 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_row_8k_9 group by c2;  
  
explain analyze select * from t_ao_row_8k_9 where c2=1;  
  
create index idx_t_ao_row_8k_9 on t_ao_row_8k_9(c2);  
  
explain analyze select * from t_ao_row_8k_9 where c2=1;  
  
-- 压缩比9, 2M  
  
create table t_ao_row_2m_9(id int, c1 text, c2 int) with (appendonly=true, blocksize=2097152, ORIENTATION=row, COMPRESSTYPE=zlib, COMPRESSLEVEL=9, CHECKSUM=false);  
  
insert into t_ao_row_2m_9 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_row_2m_9 group by c2;  
  
explain analyze select * from t_ao_row_2m_9 where c2=1;  
  
create index idx_t_ao_row_2m_9 on t_ao_row_2m_9(c2);  
  
explain analyze select * from t_ao_row_2m_9 where c2=1;  
  
-- AO 列存  
  
-- 不压缩, 8K  
  
create table t_ao_COLUMN_8k_0(id int, c1 text, c2 int) with (appendonly=true, blocksize=8192, ORIENTATION=COLUMN, COMPRESSTYPE=NONE, CHECKSUM=false);  
  
insert into t_ao_COLUMN_8k_0 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_COLUMN_8k_0 group by c2;  
  
explain analyze select * from t_ao_COLUMN_8k_0 where c2=1;  
  
create index idx_t_ao_COLUMN_8k_0 on t_ao_COLUMN_8k_0(c2);  
  
explain analyze select * from t_ao_COLUMN_8k_0 where c2=1;  
  
-- 不压缩, 2M  
  
create table t_ao_COLUMN_2m_0(id int, c1 text, c2 int) with (appendonly=true, blocksize=2097152, ORIENTATION=COLUMN, COMPRESSTYPE=NONE, CHECKSUM=false);  
  
insert into t_ao_COLUMN_2m_0 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_COLUMN_2m_0 group by c2;  
  
explain analyze select * from t_ao_COLUMN_2m_0 where c2=1;  
  
create index idx_t_ao_COLUMN_2m_0 on t_ao_COLUMN_2m_0(c2);  
  
explain analyze select * from t_ao_COLUMN_2m_0 where c2=1;  
  
-- 压缩比5, 8K  
  
create table t_ao_COLUMN_8k_5(id int, c1 text, c2 int) with (appendonly=true, blocksize=8192, ORIENTATION=COLUMN, COMPRESSTYPE=zlib, COMPRESSLEVEL=5, CHECKSUM=false);  
  
insert into t_ao_COLUMN_8k_5 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_COLUMN_8k_5 group by c2;  
  
explain analyze select * from t_ao_COLUMN_8k_5 where c2=1;  
  
create index idx_t_ao_COLUMN_8k_5 on t_ao_COLUMN_8k_5(c2);  
  
explain analyze select * from t_ao_COLUMN_8k_5 where c2=1;  
  
-- 压缩比5, 2M  
  
create table t_ao_COLUMN_2m_5(id int, c1 text, c2 int) with (appendonly=true, blocksize=2097152, ORIENTATION=COLUMN, COMPRESSTYPE=zlib, COMPRESSLEVEL=5, CHECKSUM=false);  
  
insert into t_ao_COLUMN_2m_5 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_COLUMN_2m_5 group by c2;  
  
explain analyze select * from t_ao_COLUMN_2m_5 where c2=1;  
  
create index idx_t_ao_COLUMN_2m_5 on t_ao_COLUMN_2m_5(c2);  
  
explain analyze select * from t_ao_COLUMN_2m_5 where c2=1;  
  
-- 压缩比9, 8K  
  
create table t_ao_COLUMN_8k_9(id int, c1 text, c2 int) with (appendonly=true, blocksize=8192, ORIENTATION=COLUMN, COMPRESSTYPE=zlib, COMPRESSLEVEL=9, CHECKSUM=false);  
  
insert into t_ao_COLUMN_8k_9 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_COLUMN_8k_9 group by c2;  
  
explain analyze select * from t_ao_COLUMN_8k_9 where c2=1;  
  
create index idx_t_ao_COLUMN_8k_9 on t_ao_COLUMN_8k_9(c2);  
  
explain analyze select * from t_ao_COLUMN_8k_9 where c2=1;  
  
-- 压缩比9, 2M  
  
create table t_ao_COLUMN_2m_9(id int, c1 text, c2 int) with (appendonly=true, blocksize=2097152, ORIENTATION=COLUMN, COMPRESSTYPE=zlib, COMPRESSLEVEL=9, CHECKSUM=false);  
  
insert into t_ao_COLUMN_2m_9 select id, repeat(md5(random()::text), 128), random()*10000 from generate_series(1,10000000) t(id);  
  
explain analyze select c2,count(*) from t_ao_COLUMN_2m_9 group by c2;  
  
explain analyze select * from t_ao_COLUMN_2m_9 where c2=1;  
  
create index idx_t_ao_COLUMN_2m_9 on t_ao_COLUMN_2m_9(c2);  
  
explain analyze select * from t_ao_COLUMN_2m_9 where c2=1;  
  
select tablename, pg_size_pretty(pg_total_relation_size('postgres.'||tablename)) from pg_tables where schemaname='postgres';  

nohup psql -f ./test.sql > ./log 2>&1 &

结果对比

存储形态 写入1000万耗时 空间占用 分组聚合 单值多行查询(走全表) 建索引耗时 单值多行查询(走索引)
堆表 120秒 44GB 1.8秒 1.5秒 13秒 37毫秒
AO行存8K不压缩 81秒 1.3GB 168毫秒 96毫秒 356毫秒 32毫秒
AO行存2MB不压缩 101秒 39GB 1.8秒 1.7秒 1.9秒 158毫秒
AO行存8K压缩5级 80秒 557MB 322毫秒 269毫秒 505毫秒 54毫秒
AO行存2MB压缩5级 104秒 690MB 1.6秒 1.5秒 1.7秒 163毫秒
AO行存8K压缩9级 80秒 557MB 331毫秒 247毫秒 500毫秒 58毫秒
AO行存2MB压缩9级 106秒 690MB 1.7秒 1.38秒 1.67秒 162毫秒
AO列存8K不压缩 92.6秒 39GB 362毫秒 622毫秒 877毫秒 36毫秒
AO列存2MB不压缩 98.8秒 38GB 140毫秒 1.62秒 1.8秒 176毫秒
AO列存8K压缩5级 83秒 1.4GB 125毫秒 2.2秒 2.5秒 58毫秒
AO列存2MB压缩5级 104秒 593MB 152毫秒 1.37秒 1.73秒 189毫秒
AO列存8K压缩9级 83秒 1.4GB 122毫秒 2.3秒 2.5秒 62毫秒
AO列存2MB压缩9级 106秒 593MB 136毫秒 1.5秒 1.77秒 181毫秒

云端产品

阿里云 RDS PostgreSQL

阿里云 HybridDB for PostgreSQL

Flag Counter

digoal’s 大量PostgreSQL文章入口

You May Also Enjoy

PostgreSQL(PPAS 兼容Oracle) 从零开始入门手册 - 珍藏版

17 minute read

背景 云数据库PPAS版，是阿里云与EnterpriseDB公司(简称EDB)合作基于PostgreSQL高度兼容Oracle语法的数据库服务，为用户提供易于操作的迁移工具，兼容范围涵盖：PL/SQL、数据类型、高级函数、表分区等。 用户可以直接在阿里云购买PPAS进行使用。 如果在购买PPAS前，想试用一下...

PostgreSQL pipelinedb 流计算插件 - IoT应用 - 实时轨迹聚合

1 minute read

背景 IoT场景，车联网场景，共享单车场景，人的行为位点等，终端实时上报的是孤立的位点，我们需要将其补齐成轨迹。 例如共享单车，下单，开锁，生成订单，骑行，关闭订单，关锁。这个过程有一个唯一的订单号，每次上报的位点会包含时间，订单号，位置。 根据订单号，将点聚合为轨迹。 使用pipelinedb插件，可以实...

PostgreSQL datediff 日期间隔（单位转换）兼容SQL用法

2 minute read

背景 使用datediff，对时间或日期相减，得到的间隔，转换为目标单位(日、月、季度、年、小时、秒。。。等)的数值。 DATEDIFF ( datepart, {date|timestamp}, {date|timestamp} ) 周 select datediff(week,'2009-01-0...