视图与冗余物理表的查询性能测试

Posted rongyongfeikai2

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了视图与冗余物理表的查询性能测试相关的知识,希望对你有一定的参考价值。

如果有30张表,他们各自有各自的自定义字段,也有一部分公有字段;需要将公有字段归一起来统一查询,一般而言有两种方式:
1.公共字段物理表 数据在入库/更新时,更新自己的表的数据,同时亦将数据更新入公共表中
2.视图 数据在入库/更新时,更新自己的表的数据,视图因为是逻辑上的表所以查询时可以查询到更新
两种方式各有优劣:
1.公共字段方式
优点:查询公共表时占优势,sql比较简单,pg在解析和优化sql时耗时都比较短;而且数据在一张表中,不需要扫描多张表
缺点:APP表和公共字段表必须同时更新,需要事务[事务由于需要写预写日志等,会降低写性能];如果表上有索引时,增更删时均需要更新索引,也会降低写性能,尤其是表不断变大的情况下
2.视图的方式
优点:只需要更新各自的APP表即可,各APP表的数据量均会小于总数据量,写占优
缺点:表数量上升,涉及到多表union all,查询时会慢
--当然以上的查询均是指用到公共字段的查询时,APP内部只用自己的APP表,所以查询性能并不受上述影响


所以这两种方式类似于一个南拳北腿,各有擅长的领域,并不好粗暴的进行比较。
不过,对于视图的查询性能差到怎样一个地步,我们还是颇有兴趣的。


环境:单机实体机,8个逻辑核,32G内存,CentOS7.0操作系统,PostgreSQL9.4
表:

APP表:tb_test_1/.../tb_test_30[总共30张表,分别有starttime,endtime以及col[表id]_1/col[表id]_2/.../col[表id]_45总共47个字段

如下图所示:


公共字段表:tb_test_common,分别有startime,endtime以及col0_1/.../col0_40(对应app表的前40个字段)总共42个字段
视图:v_evt,分别有startime,endtime以及col0_1/.../col0_40(对应app表的前40个字段)总共42个字段
APP表每张表100万数据量,30个APP表,对应到tb_test_common中总共3000万数据量


在每张表的starttime和endtime上均建立索引,增快按照starttime和endtime进行where查询时的速度【当然,前面提过,索引一方面消耗存储空间,一方面insert/update时均需要更新索引,降低写速度】

我们查一下执行规划,可以清晰看到select语句是走了bitmap index scan的。



数据准备好了,我们设计了5种比对方式,分别为:






执行结果(第一次执行结果,后面视图查询速度会明显加快,因为有执行规划的缓存)


我用个表格展示,会更加清晰:

 

公共字段表

视图

where条件过滤

0.0371432304382

0.379799127579

过滤条件+order by limit 10

11.1989729404

44.5428109169

聚合函数

50.8744649887

50.5313501358

不进行过滤,全表order by取limit

0.122601985931

0.368887901306

count

43.4130840302

67.3434300423


上面的时间均是以秒作为单位。大家应该对性能大概有了个感性的认识。


====华丽分割线,以下是构建数据的脚本======================

#coding:utf-8
import psycopg2
import traceback
import random
import time

HOST="localhost"
DBNAME=""
USERNAME=""
PASSWORD=""

#数据库时间为2016-05-23当天
starttime = 1463932800
endtime = 1464019199

def create_tables(nums):
    arr = list()
    for i in range(0,nums):
        tablename = "tb_test_"+str(i)
        if i == 0:
            tablename = 'tb_test_common'
        sql = "create table " + tablename + "("\\
            + "starttime integer,"\\
            + "endtime integer,"
        colnum = 46
        #公共表时间字段+40个字段,普通APP表时间字段+45个字段
        if i == 0:
            colnum = 41
        for j in range(1,colnum):
            sql += "col"+str(i)+"_"+str(j)+" varchar(1024)"
            if j != colnum - 1:
                sql += ","
        sql += ")"
        arr.append(sql)

    return ";".join(arr)

#写一个一百万数据的文件
def write_million_data(filepath):
    print "create data file start ..."
    f = open(filepath,"w")
    for i in range(0,1000000):
        curstart = random.randint(starttime,endtime)
        f.write(str(curstart)+",")
        f.write(str(random.randint(curstart,endtime))+",")
        for j in range(0,45):
            f.write("hello")
            if j != 44:
                f.write(",")
        f.write("\\n")
    f.close()
    print "create data file end ..."

def create_view(nums):
    sql = "create view v_evt as "
    for i in range(1,nums):
        sql += "select starttime,endtime,"
        for j in range(1,41):
            sql += "col"+str(i)+"_"+str(j)+" as col_0_"+str(j)
            if j != 40:
                sql += ","
        sql += " from tb_test_"+str(i)+" "
        if i != nums-1:
            sql += " union all "
    return sql

def init_all():
    filepath = "/tmp/view_data"
    write_million_data(filepath)
    conn = psycopg2.connect(host=HOST,dbname=DBNAME,user=USERNAME,password=PASSWORD)
    cur = conn.cursor()
    #新建31张表;第一张为公有字段表,其他30张为app表
    cur.execute(create_tables(31))
    conn.commit()
    print "create tables"
    #将数据copy入第1-30的APP表和公共表
    for i in range(1,31):
        cur.copy_from(open(filepath,"r"),"tb_test_"+str(i),sep=",")
        print "copy data into table"
    conn.commit()
    for i in range(1,31):
        insql = "insert into tb_test_common select starttime,endtime,"
        for j in range(1,41):
            insql += "col"+str(i)+"_"+str(j)
            if j != 40:
                insql += ','
        insql += " from tb_test_"+str(i)
        print insql
        cur.execute(insql)
    conn.commit()
    #在starttime和endtime字段建立索引
    for i in range(1,31):
        cur.execute("create index index_tb_test_"+str(i)+"_starttime on tb_test_"+str(i)+"(starttime)")
        cur.execute("create index index_tb_test_"+str(i)+"_endime on tb_test_"+str(i)+"(endtime)")
        conn.commit()
        print "create index on index_tb_test_"+str(i)
    cur.execute("create index index_tb_test_common_starttime on tb_test_common(starttime)")
    cur.execute("create index index_tb_test_common_endime on tb_test_common(endtime)")
    conn.commit()
    print "create index on tb_test_common"
    #建立视图
    vsql = create_view(31)
    cur.execute(vsql)
    conn.commit()
    print "create view"

    #将数据写入40张表
    conn.close()

if __name__ == '__main__':
    #init_all()

    '''冗余表的优点在于查询,而缺点在于增删改[需要事务,同时需要写两份]'''
    '''视图的优点在于增删改只需要改原始物理表,而缺点在于查询'''
    '''所以以下测试是用冗余表之长对决视图之短'''

    '''为了查询性能非常优秀,我为公共表和普通表的字段startime和endtime字段建立索引,并在查询条件中使用它'''
    '''但众所周知的,有索引意味着insert和update均需要更新索引,所以写入速度是影响的'''
    
    #进行性能测试
    conn = psycopg2.connect(host=HOST,dbname=DBNAME,user=USERNAME,password=PASSWORD)
    cur = conn.cursor()

    #第零组,验证 where条件过滤 的速度
    stime = time.time()
    cur.execute("select count(*) from tb_test_common where starttime=1463932800")
    etime = time.time()
    print "query tb_test_common, waste time:",etime-stime,"s;count:",cur.fetchone()
    
    stime = time.time()
    cur.execute("select count(*) from v_evt where starttime=1463932800")
    etime = time.time()
    print "query v_evt, waste time:",etime-stime,"s;count:",cur.fetchone()
    
    print "*************************************************************"

    #第一组,验证 过滤条件+order by limit 10 的速度
    stime = time.time()
    cur.execute("select * from tb_test_common where starttime>=1463932800 and starttime<=1463936400 order by starttime desc")
    etime = time.time()
    print "query tb_test_common, waste time:",etime-stime,"s;count:",len(cur.fetchall())

    stime = time.time()
    cur.execute("select * from v_evt where starttime>=1463932800 and starttime<=1463936400 order by starttime desc")
    etime = time.time()
    print "query v_event, waste time:",etime-stime,"s;count:",len(cur.fetchall())
    
    print "*************************************************************"

    #第二组,验证 聚合函数 的速度
    stime = time.time()
    cur.execute("select to_char(to_timestamp(starttime),'YYYY-MM-DD') as mydate,count(*) as num from tb_test_common where starttime>=1463932800 and starttime<=1463940000 group by mydate order by mydate desc")
    etime = time.time()
    print "query tb_test_common, waste time:",etime-stime,"s;count:",len(cur.fetchall())

    stime = time.time()
    cur.execute("select to_char(to_timestamp(starttime),'YYYY-MM-DD') as mydate,count(*) as num from v_evt where starttime>=1463932800 and starttime<=1463940000 group by mydate order by mydate desc")
    etime = time.time()
    print "query v_event, waste time:",etime-stime,"s;count:",len(cur.fetchall())
    print "*************************************************************"

    #第三组,验证 不进行过滤,全表order by 返回limit的速度
    stime = time.time()
    cur.execute("select * from tb_test_common order by starttime desc limit 100")
    etime = time.time()
    print "query tb_test_common, waste time:",etime-stime,"s;count:",len(cur.fetchall())


    stime = time.time()
    cur.execute("select * from v_evt order by starttime desc limit 100")
    etime = time.time()
    print "query v_event, waste time:",etime-stime,"s;count:",len(cur.fetchall())
    print "*************************************************************"

    #第四组,验证 count 的速度
    stime = time.time()
    cur.execute("select count(*) from tb_test_common")
    etime = time.time()
    print "query tb_test_common, waste time:",etime-stime,"s;count:",cur.fetchone()

    stime = time.time()
    cur.execute("select count(*) from v_evt")
    etime = time.time()
    print "query v_event, waste time:",etime-stime,"s;count:",cur.fetchone()

    cur.close()
    conn.close()


另外附一张二次查询的结果,可以看到视图的查询性能提升明显(因为查询规划被缓存了)



以上是关于视图与冗余物理表的查询性能测试的主要内容,如果未能解决你的问题,请参考以下文章

与查询相比的视图性能(使用联合语句)

CentOS 7中RAID5搭建与测试

优雅与性能:何时使用冗余列表进行查询?

未分类

使用从循环内的代码片段中提取的函数避免代码冗余/计算开销

SQL执行效率和性能测试方法(SqlServer)