基于PostgreSQL的流式PipelineDB, 1000万/s实时统计不是梦
背景
PipelineDB是基于PostgreSQL研发的一种流式关系数据库(0.8.1基于9.4.4)，这种数据库的特点是自动处理流式数据，不存储原始数据，只存储处理后的数据，所以非常适合当下流行的实时流式数据处理。
例如网站流量统计，IT服务的监控统计，APPStore的浏览统计，物联网的传感器数据实时统计，物流订单的实时统计等等。
pipelineDB介绍
http://www.postgresql.org/about/news/1596/
PipelineDB, an open-source relational streaming-SQL database, publicly released version (0.7.7) today and made the product available as open-source via their website and GitHub.
PipelineDB is based on, and is wire compatible with, PostgreSQL 9.4 and has added functionality including continuous SQL queries, probabilistic data structures, sliding windowing, and stream-table joins.
For a full description of PipelineDB and its capabilities see their technical documentation.
PipelineDB’s fundamental abstraction is what is called a continuous view.
These are much like regular SQL views, except that their defining SELECT queries can include streams as a source to read from.
The most important property of continuous views is that they only store their output in the database.
That output is then continuously updated incrementally as new data flows through streams, and raw stream data is discarded once all continuous views have read it.
Let's look at a canonical example:
CREATE CONTINUOUS VIEW v AS SELECT COUNT(*) FROM stream
Only one row would ever physically exist in PipelineDB for this continuous view, and its value would simply be incremented for each new event ingested.
For more information on PipelineDB as a company, product and for examples and benefits, please check out their first blog post on their new website.
流式统计例子
创建动态流视图，不需要对表进行定义，太棒了，这类似活生生的NoSQL。
pipeline=# CREATE CONTINUOUS VIEW v0 AS SELECT COUNT(*) FROM stream;
CREATE CONTINUOUS VIEW
pipeline=# CREATE CONTINUOUS VIEW v1 AS SELECT COUNT(*) FROM stream;
CREATE CONTINUOUS VIEW
激活流视图
pipeline=# ACTIVATE;
ACTIVATE 2
往流写入数据
pipeline=# INSERT INTO stream (x) VALUES (1);
INSERT 0 1
pipeline=# SET stream_targets TO v0;
SET
pipeline=# INSERT INTO stream (x) VALUES (1);
INSERT 0 1
pipeline=# SET stream_targets TO DEFAULT;
SET
pipeline=# INSERT INTO stream (x) VALUES (1);
INSERT 0 1
如果不想接收流数据了，停止即可
pipeline=# DEACTIVATE;
DEACTIVATE 2
查询流视图
pipeline=# SELECT count FROM v0;
count
-------
3
(1 row)
pipeline=# SELECT count FROM v1;
count
-------
2
(1 row)
pipeline=#
如何部署
安装
[root@digoal soft_bak]# rpm -ivh pipelinedb-0.8.1-centos6-x86_64.rpm
Preparing... ########################################### [100%]
1:pipelinedb ########################################### [100%]
/sbin/ldconfig: /opt/gcc4.9.3/lib/libstdc++.so.6.0.20-gdb.py is not an ELF file - it has the wrong magic bytes at the start.
/sbin/ldconfig: /opt/gcc4.9.3/lib64/libstdc++.so.6.0.20-gdb.py is not an ELF file - it has the wrong magic bytes at the start.
____ _ ___ ____ ____
/ __ \(_)___ ___ / (_)___ ___ / __ \/ __ )
/ /_/ / / __ \/ _ \/ / / __ \/ _ \/ / / / __ |
/ ____/ / /_/ / __/ / / / / / __/ /_/ / /_/ /
/_/ /_/ .___/\___/_/_/_/ /_/\___/_____/_____/
/_/
PipelineDB successfully installed. To get started, initialize a
database directory:
pipeline-init -D <data directory>
where <data directory> is a nonexistent directory where you'd
like all of your database files to live.
You can find the PipelineDB documentation at:
http://docs.pipelinedb.com
配置
[root@digoal soft_bak]# cd /usr/lib/pipelinedb
[root@digoal pipelinedb]# ll
total 16
drwxr-xr-x 2 root root 4096 Oct 15 10:47 bin
drwxr-xr-x 5 root root 4096 Oct 15 10:47 include
drwxr-xr-x 6 root root 4096 Oct 15 10:47 lib
drwxr-xr-x 4 root root 4096 Oct 15 10:47 share
[root@digoal pipelinedb]# useradd pdb
[root@digoal pipelinedb]# vi /home/pdb/.bash_profile
# add by digoal
export PS1="$USER@`/bin/hostname -s`-> "
export PGPORT=1953
export PGDATA=/data01/pg_root_1953
export LANG=en_US.utf8
export PGHOME=/usr/lib/pipelinedb
export LD_LIBRARY_PATH=$PGHOME/lib:/lib64:/usr/lib64:/usr/local/lib64:/lib:/usr/lib:/usr/local/lib:$LD_LIBRARY_PATH
export DATE=`date +"%Y%m%d%H%M"`
export PATH=$PGHOME/bin:$PATH:.
export MANPATH=$PGHOME/share/man:$MANPATH
export PGHOST=$PGDATA
export PGDATABASE=pipeline
export PGUSER=postgres
alias rm='rm -i'
alias ll='ls -lh'
unalias vi
[root@digoal pipelinedb]# mkdir /data01/pg_root_1953
[root@digoal pipelinedb]# chown pdb:pdb /data01/pg_root_1953
[root@digoal pipelinedb]# chmod 700 /data01/pg_root_1953
[root@digoal pipelinedb]# su - pdb
pdb@digoal-> which psql
/usr/lib/pipelinedb/bin/psql
初始化数据库
pdb@digoal-> psql -V
psql (PostgreSQL) 9.4.4
pdb@digoal-> cd /usr/lib/pipelinedb/bin/
pdb@digoal-> ll
total 13M
-rwxr-xr-x 1 root root 62K Sep 18 01:01 clusterdb
-rwxr-xr-x 1 root root 62K Sep 18 01:01 createdb
-rwxr-xr-x 1 root root 66K Sep 18 01:01 createlang
-rwxr-xr-x 1 root root 63K Sep 18 01:01 createuser
-rwxr-xr-x 1 root root 44K Sep 18 01:02 cs2cs
-rwxr-xr-x 1 root root 58K Sep 18 01:01 dropdb
-rwxr-xr-x 1 root root 66K Sep 18 01:01 droplang
-rwxr-xr-x 1 root root 58K Sep 18 01:01 dropuser
-rwxr-xr-x 1 root root 776K Sep 18 01:01 ecpg
-rwxr-xr-x 1 root root 28K Sep 18 00:57 gdaladdo
-rwxr-xr-x 1 root root 79K Sep 18 00:57 gdalbuildvrt
-rwxr-xr-x 1 root root 1.3K Sep 18 00:57 gdal-config
-rwxr-xr-x 1 root root 33K Sep 18 00:57 gdal_contour
-rwxr-xr-x 1 root root 188K Sep 18 00:57 gdaldem
-rwxr-xr-x 1 root root 74K Sep 18 00:57 gdalenhance
-rwxr-xr-x 1 root root 131K Sep 18 00:57 gdal_grid
-rwxr-xr-x 1 root root 83K Sep 18 00:57 gdalinfo
-rwxr-xr-x 1 root root 90K Sep 18 00:57 gdallocationinfo
-rwxr-xr-x 1 root root 42K Sep 18 00:57 gdalmanage
-rwxr-xr-x 1 root root 236K Sep 18 00:57 gdal_rasterize
-rwxr-xr-x 1 root root 25K Sep 18 00:57 gdalserver
-rwxr-xr-x 1 root root 77K Sep 18 00:57 gdalsrsinfo
-rwxr-xr-x 1 root root 49K Sep 18 00:57 gdaltindex
-rwxr-xr-x 1 root root 33K Sep 18 00:57 gdaltransform
-rwxr-xr-x 1 root root 158K Sep 18 00:57 gdal_translate
-rwxr-xr-x 1 root root 168K Sep 18 00:57 gdalwarp
-rwxr-xr-x 1 root root 41K Sep 18 01:02 geod
-rwxr-xr-x 1 root root 1.3K Sep 18 00:51 geos-config
lrwxrwxrwx 1 root root 4 Oct 15 10:47 invgeod -> geod
lrwxrwxrwx 1 root root 4 Oct 15 10:47 invproj -> proj
-rwxr-xr-x 1 root root 20K Sep 18 01:02 nad2bin
-rwxr-xr-x 1 root root 186K Sep 18 00:57 nearblack
-rwxr-xr-x 1 root root 374K Sep 18 00:57 ogr2ogr
-rwxr-xr-x 1 root root 77K Sep 18 00:57 ogrinfo
-rwxr-xr-x 1 root root 283K Sep 18 00:57 ogrlineref
-rwxr-xr-x 1 root root 47K Sep 18 00:57 ogrtindex
-rwxr-xr-x 1 root root 30K Sep 18 01:01 pg_config
-rwxr-xr-x 1 root root 30K Sep 18 01:01 pg_controldata
-rwxr-xr-x 1 root root 33K Sep 18 01:01 pg_isready
-rwxr-xr-x 1 root root 39K Sep 18 01:01 pg_resetxlog
-rwxr-xr-x 1 root root 183K Sep 18 01:02 pgsql2shp
lrwxrwxrwx 1 root root 4 Oct 15 10:47 pipeline -> psql
-rwxr-xr-x 1 root root 74K Sep 18 01:01 pipeline-basebackup
lrwxrwxrwx 1 root root 9 Oct 15 10:47 pipeline-config -> pg_config
-rwxr-xr-x 1 root root 44K Sep 18 01:01 pipeline-ctl
-rwxr-xr-x 1 root root 355K Sep 18 01:01 pipeline-dump
-rwxr-xr-x 1 root root 83K Sep 18 01:01 pipeline-dumpall
-rwxr-xr-x 1 root root 105K Sep 18 01:01 pipeline-init
-rwxr-xr-x 1 root root 50K Sep 18 01:01 pipeline-receivexlog
-rwxr-xr-x 1 root root 56K Sep 18 01:01 pipeline-recvlogical
-rwxr-xr-x 1 root root 153K Sep 18 01:01 pipeline-restore
-rwxr-xr-x 1 root root 6.2M Sep 18 01:01 pipeline-server
lrwxrwxrwx 1 root root 15 Oct 15 10:47 postmaster -> pipeline-server
-rwxr-xr-x 1 root root 49K Sep 18 01:02 proj
-rwxr-xr-x 1 root root 445K Sep 18 01:01 psql
-rwxr-xr-x 1 root root 439K Sep 18 01:02 raster2pgsql
-rwxr-xr-x 1 root root 62K Sep 18 01:01 reindexdb
-rwxr-xr-x 1 root root 181K Sep 18 01:02 shp2pgsql
-rwxr-xr-x 1 root root 27K Sep 18 00:57 testepsg
-rwxr-xr-x 1 root root 63K Sep 18 01:01 vacuumdb
pdb@digoal-> pipeline-init -D $PGDATA -U postgres -E UTF8 --locale=C -W
pdb@digoal-> cd $PGDATA
pdb@digoal-> ll
total 108K
drwx------ 5 pdb pdb 4.0K Oct 15 10:57 base
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 global
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_clog
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_dynshmem
-rw------- 1 pdb pdb 4.4K Oct 15 10:57 pg_hba.conf
-rw------- 1 pdb pdb 1.6K Oct 15 10:57 pg_ident.conf
drwx------ 4 pdb pdb 4.0K Oct 15 10:57 pg_logical
drwx------ 4 pdb pdb 4.0K Oct 15 10:57 pg_multixact
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_notify
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_replslot
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_serial
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_snapshots
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_stat
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_stat_tmp
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_subtrans
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_tblspc
drwx------ 2 pdb pdb 4.0K Oct 15 10:57 pg_twophase
-rw------- 1 pdb pdb 4 Oct 15 10:57 PG_VERSION
drwx------ 3 pdb pdb 4.0K Oct 15 10:57 pg_xlog
-rw------- 1 pdb pdb 88 Oct 15 10:57 pipelinedb.auto.conf
-rw------- 1 pdb pdb 23K Oct 15 10:57 pipelinedb.conf
和流处理相关的参数，例如设置内存大小，是否同步，合并的batch，工作进程数等等。
pipelinedb.conf
#------------------------------------------------------------------------------
# CONTINUOUS VIEW OPTIONS
#------------------------------------------------------------------------------
# size of the buffer for storing unread stream tuples
#tuple_buffer_blocks = 128MB
# synchronization level for combiner commits; off, local, remote_write, or on
#continuous_query_combiner_synchronous_commit = off
# maximum amount of memory to use for combiner query executions
#continuous_query_combiner_work_mem = 256MB
# maximum memory to be used by the combiner for caching; this is independent
# of combiner_work_mem
#continuous_query_combiner_cache_mem = 32MB
# the default fillfactor to use for continuous views
#continuous_view_fillfactor = 50
# the time in milliseconds a continuous query process will wait for a batch
# to accumulate
# continuous_query_max_wait = 10
# the maximum number of events to accumulate before executing a continuous query
# plan on them
#continuous_query_batch_size = 10000
# the number of parallel continuous query combiner processes to use for
# each database
#continuous_query_num_combiners = 2
# the number of parallel continuous query worker processes to use for
# each database
#continuous_query_num_workers = 2
# allow direct changes to be made to materialization tables?
#continuous_query_materialization_table_updatable = off
# inserts into streams should be synchronous?
#synchronous_stream_insert = off
# continuous views that should be affected when writing to streams.
# it is string with comma separated values for continuous view names.
#stream_targets = ''
启动数据库，可以看到原生是支持postgis的，吐个槽，这个项目是专门为NASA研发的么？
pdb@digoal-> pipeline-ctl start
pdb@digoal-> psql pipeline postgres
psql (9.4.4)
Type "help" for help.
pipeline=# \l
List of databases
Name | Owner | Encoding | Collate | Ctype | Access privileges
-----------+----------+----------+---------+-------+-----------------------
pipeline | postgres | UTF8 | C | C |
template0 | postgres | UTF8 | C | C | =c/postgres +
| | | | | postgres=CTc/postgres
template1 | postgres | UTF8 | C | C | =c/postgres +
| | | | | postgres=CTc/postgres
(3 rows)
pipeline=# \dx
List of installed extensions
Name | Version | Schema | Description
------------------+----------+------------+---------------------------------------------------------------------
plpgsql | 1.0 | pg_catalog | PL/pgSQL procedural language
postgis | 2.2.0dev | pg_catalog | PostGIS geometry, geography, and raster spatial types and functions
postgis_topology | 2.2.0dev | topology | PostGIS topology spatial types and functions
(3 rows)
查看pipelinedb加了哪些函数，有些是插件形式加入的，如POSTGIS，有些是我们可以借鉴，直接拿来用的。
pipeline=# select proname from pg_proc order by oid desc;
......
second
minute
hour
day
month
year
......
cmsketch_empty
tdigest_add
tdigest_empty
tdigest_empty
bloom_add
bloom_empty
bloom_empty
hll_add
hll_empty
hll_empty
......
在我自己的笔记本中的虚拟机中的性能测试
创建5个动态流视图，动态流视图就是不需要建立基表的流视图。
CREATE CONTINUOUS VIEW v0 AS SELECT COUNT(*) FROM stream;
CREATE CONTINUOUS VIEW v1 AS SELECT sum(x::int),count(*),avg(y::int) FROM stream;
CREATE CONTINUOUS VIEW v001 AS SELECT sum(x::int),count(*),avg(y::int) FROM stream1;
CREATE CONTINUOUS VIEW v002 AS SELECT sum(x::int),count(*),avg(y::int) FROM stream2;
CREATE CONTINUOUS VIEW v003 AS SELECT sum(x::int),count(*),avg(y::int) FROM stream3;
激活流统计
activate;
查看数据字典
select relname from pg_class where relkind='C';
批量插入测试
pdb@digoal-> vi test.sql
insert into stream(x,y,z) select generate_series(1,1000),1,1;
insert into stream1(x,y,z) select generate_series(1,1000),1,1;
insert into stream2(x,y,z) select generate_series(1,1000),1,1;
insert into stream3(x,y,z) select generate_series(1,1000),1,1;
测试结果，注意这里需要使用simple或者extended ， 如果用prepared会导致只有最后一条SQL起作用。现在不清楚是pipelinedb还是pgbench的BUG。
pdb@digoal-> /opt/pgsql/bin/pgbench -M extended -n -r -f ./test.sql -P 1 -c 10 -j 10 -T 100000
progress: 1.0 s, 133.8 tps, lat 68.279 ms stddev 58.444
progress: 2.0 s, 143.9 tps, lat 71.623 ms stddev 53.880
progress: 3.0 s, 149.5 tps, lat 66.452 ms stddev 49.727
progress: 4.0 s, 148.3 tps, lat 67.085 ms stddev 55.484
progress: 5.1 s, 145.7 tps, lat 68.624 ms stddev 67.795
每秒入库约58万条记录，并完成5个流视图的统计。
因为都在内存中完成，所以速度非常快。
pipelinedb使用了worker进程来处理数据合并。
压测时的top如下：
top - 11:23:07 up 2:49, 4 users, load average: 1.83, 3.08, 1.78
Tasks: 177 total, 5 running, 172 sleeping, 0 stopped, 0 zombie
Cpu(s): 11.6%us, 15.0%sy, 10.3%ni, 63.0%id, 0.0%wa, 0.0%hi, 0.1%si, 0.0%st
Mem: 3916744k total, 605084k used, 3311660k free, 27872k buffers
Swap: 1048572k total, 0k used, 1048572k free, 401748k cached
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
11469 pdb 25 5 405m 75m 67m R 52.9 2.0 1:56.45 pipeline: bgworker: worker0 [pipeline]
12246 pdb 20 0 400m 69m 67m S 14.3 1.8 0:10.55 pipeline: postgres pipeline [local] idle
12243 pdb 20 0 400m 69m 67m S 13.3 1.8 0:10.45 pipeline: postgres pipeline [local] idle
12248 pdb 20 0 400m 69m 67m S 13.3 1.8 0:10.40 pipeline: postgres pipeline [local] idle
12244 pdb 20 0 400m 69m 67m S 12.6 1.8 0:10.50 pipeline: postgres pipeline [local] idle
12237 pdb 20 0 400m 69m 67m R 12.3 1.8 0:10.52 pipeline: postgres pipeline [local] idle
12247 pdb 20 0 402m 70m 67m R 12.3 1.8 0:10.70 pipeline: postgres pipeline [local] idle
12245 pdb 20 0 401m 69m 67m S 12.0 1.8 0:10.78 pipeline: postgres pipeline [local] idle
12235 pdb 20 0 400m 69m 67m S 11.3 1.8 0:10.88 pipeline: postgres pipeline [local] idle
12239 pdb 20 0 400m 69m 67m S 11.0 1.8 0:10.79 pipeline: postgres pipeline [local] idle
12241 pdb 20 0 400m 69m 67m S 11.0 1.8 0:10.53 pipeline: postgres pipeline [local] idle
11466 pdb 20 0 119m 1480 908 R 5.3 0.0 0:58.39 pipeline: stats collector process
11468 pdb 25 5 401m 12m 9744 S 2.3 0.3 0:16.49 pipeline: bgworker: combiner0 [pipeline]
12228 pdb 20 0 678m 3408 884 S 2.3 0.1 0:02.36 /opt/pgsql/bin/pgbench -M extended -n -r -f ./test.sql -P 1 -c 10 -j 10 -T 100000
11464 pdb 20 0 398m 17m 16m S 1.7 0.4 0:10.47 pipeline: wal writer process
11459 pdb 20 0 398m 153m 153m S 0.0 4.0 0:00.37 /usr/lib/pipelinedb/bin/pipeline-server
11460 pdb 20 0 115m 852 424 S 0.0 0.0 0:00.02 pipeline: logger process
11462 pdb 20 0 398m 3336 2816 S 0.0 0.1 0:00.06 pipeline: checkpointer process
11463 pdb 20 0 398m 2080 1604 S 0.0 0.1 0:00.08 pipeline: writer process
11465 pdb 20 0 401m 4460 1184 S 0.0 0.1 0:00.33 pipeline: autovacuum launcher process
11467 pdb 20 0 398m 1992 1056 S 0.0 0.1 0:00.00 pipeline: continuous query scheduler process
pdb@digoal-> psql
psql (9.4.4)
Type "help" for help.
pipeline=# select * from v0;
count
---------
9732439
(1 row)
pipeline=# select * from v1;
sum | count | avg
------------+---------+------------------------
4923514276 | 9837585 | 1.00000000000000000000
(1 row)
pipeline=# select * from v001;
sum | count | avg
--------------+----------+------------------------
505023543131 | 11036501 | 1.00000000000000000000
(1 row)
pipeline=# select * from v002;
sum | count | avg
---------------+----------+------------------------
1005065536319 | 12119513 | 1.00000000000000000000
(1 row)
pipeline=# select * from v003;
sum | count | avg
-------------+----------+------------------------
14948355485 | 29867002 | 1.00000000000000000000
(1 row)
在写入 10 亿 流数据后，数据库的大小依旧只有13MB，因为流数据都在内存中，处理完就丢弃了。
pipeline=# \l+
List of databases
Name | Owner | Encoding | Collate | Ctype | Access privileges | Size | Tablespace | Description
-----------+----------+----------+---------+-------+-----------------------+-------+------------+--------------------------------------------
pipeline | postgres | UTF8 | C | C | | 13 MB | pg_default | default administrative connection database
template0 | postgres | UTF8 | C | C | =c/postgres +| 12 MB | pg_default | unmodifiable empty database
| | | | | postgres=CTc/postgres | | |
template1 | postgres | UTF8 | C | C | =c/postgres +| 12 MB | pg_default | default template for new databases
| | | | | postgres=CTc/postgres | | |
(3 rows)
如果你的应用有类似场景，恭喜你，找到终极杀手锏了。
补充物理机的测试数据
E5-2650上的测试结果。
开10个PipelineDB实例，跑以上CASE，每秒约处理600万记录。
结合lvs,haproxy或者jdbc LB，你可以搞比较大型的实时处理了，当然这也说明了PipelineDB的单机性能还有非常多的提升空间。
老版本不能充分利用CPU，新版本已经不需要开多个实例了，一个实例就可以跑满整个CPU。
参考
https://github.com/pipelinedb/pipelinedb
https://www.pipelinedb.com/