HAWQ 与 Hive 查询性能对比测试

1. 四台 VMware 虚机组成的 Hadoop 集群。
2. 每台机器配置如下：
（1）15K RPM SAS 100GB
（2）Intel(R) Xeon(R) E5-2620 v2 @ 2.10GHz，双核双 CPU
（3）8GB 内存，8GB Swap
（4）10000 Mb/s 虚拟网卡

三、软件环境

1. Linux：CentOS release 6.4，核心2.6.32-358.el6.x86_64
2. Ambari：2.4.1
3. Hadoop：HDP 2.5.0
4. Hive（Hive on Tez）：2.1.0
5. HAWQ：2.1.1.0
6. HAWQ PXF：3.1.1

四、数据模型

1. 表结构

实验模拟一个记录页面点击数据的应用场景。数据模型中包含日期、页面、浏览器、引用、状态 5 个维度表，1 个页面点击事实表。表结构和关系如图1 所示。

图1

2. 记录数

各表的记录数如表1 所示。

表名	行数
page_click_fact	1亿
page_dim	20万
referrer_dim	100万
browser_dim	2万
status_code	70
date_dim	366

表1

五、建表并生成数据

1. 建立 hive 库表

create database test;
use test;

create table browser_dim(
  browser_sk bigint, 
  browser_nm varchar(100), 
  browser_version_no varchar(100), 
  flash_version_no varchar(100), 
  flash_enabled_flg int, 
  java_version_no varchar(100), 
  platform_desc string, 
  java_enabled_flg int, 
  java_script_enabled_flg int, 
  cookies_enabled_flg int, 
  user_language_cd varchar(100), 
  screen_color_depth_no varchar(100), 
  screen_size_txt string)
row format delimited 
  fields terminated by ',' 
stored as orc; 

create table date_dim(
  cal_dt date, 
  day_in_cal_yr_no int, 
  day_of_week_no int, 
  start_of_month_dt date, 
  start_of_quarter_dt date, 
  start_of_week_dt date, 
  start_of_year_dt date)
row format delimited 
  fields terminated by ',' 
stored as orc;

create table page_dim(
  page_sk bigint, 
  domain_nm varchar(200), 
  reachability_cd string, 
  page_desc string, 
  protocol_nm varchar(20))
row format delimited 
  fields terminated by ',' 
stored as orc;
 
create table referrer_dim(
  referrer_sk bigint, 
  referrer_txt string, 
  referrer_domain_nm varchar(200))
row format delimited 
  fields terminated by ',' 
stored as orc;

create table status_code_dim(
  status_cd varchar(100), 
  client_error_flg int, 
  status_cd_desc string, 
  server_error_flg int)
row format delimited 
  fields terminated by ',' 
stored as orc;

create table page_click_fact(
  visitor_id varchar(100), 
  detail_tm timestamp, 
  page_click_dt date, 
  page_sk bigint, 
  client_session_dt date, 
  previous_page_sk bigint, 
  referrer_sk bigint, 
  next_page_sk bigint, 
  status_cd varchar(100), 
  browser_sk bigint, 
  bytes_received_cnt bigint, 
  bytes_sent_cnt bigint, 
  client_detail_tm timestamp, 
  entry_point_flg int, 
  exit_point_flg int, 
  ip_address varchar(20), 
  query_string_txt string, 
  seconds_spent_on_page_cnt int, 
  sequence_no int, 
  requested_file_txt string)
row format delimited 
  fields terminated by ',' 
stored as orc;

说明：hive 表使用 ORCfile 存储格式。

2. 用 Java 程序生成 hive 表数据

ORC 压缩后的各表对应的 HDFS 文件大小如下：

2.2 M   /apps/hive/warehouse/test.db/browser_dim
641     /apps/hive/warehouse/test.db/date_dim
4.1 G   /apps/hive/warehouse/test.db/page_click_fact
16.1 M  /apps/hive/warehouse/test.db/page_dim
22.0 M  /apps/hive/warehouse/test.db/referrer_dim
1.1 K   /apps/hive/warehouse/test.db/status_code_dim

3. 分析 hive 表

analyze table date_dim compute statistics;
analyze table browser_dim compute statistics;
analyze table page_dim compute statistics;
analyze table referrer_dim compute statistics;
analyze table status_code_dim compute statistics;
analyze table page_click_fact compute statistics;

4. 建立 HAWQ 外部表

create schema ext;
set search_path=ext;

create external table date_dim(
cal_dt              date,                                    
day_in_cal_yr_no    int4,                                     
day_of_week_no      int4,                                     
start_of_month_dt   date,                                    
start_of_quarter_dt date,                                    
start_of_week_dt    date,                                    
start_of_year_dt    date  
)
location ('pxf://hdp1:51200/test.date_dim?profile=hiveorc')  
format 'custom' (formatter='pxfwritable_import'); 

create external table browser_dim(
browser_sk              int8,                                  
browser_nm              varchar(100),                            
browser_version_no      varchar(100),                            
flash_version_no        varchar(100),                            
flash_enabled_flg       int,                                     
java_version_no         varchar(100),                            
platform_desc           text,                                  
java_enabled_flg        int,    
java_script_enabled_flg int,    
cookies_enabled_flg     int,                                     
user_language_cd        varchar(100),         
screen_color_depth_no   varchar(100),            
screen_size_txt         text   
)  
location ('pxf://hdp1:51200/test.browser_dim?profile=hiveorc')  
format 'custom' (formatter='pxfwritable_import'); 

create external table page_dim(
page_sk             int8,                                  
domain_nm           varchar(200),                            
reachability_cd     text,                           
page_desc           text,                                  
protocol_nm         varchar(20) 
) 
location ('pxf://hdp1:51200/test.page_dim?profile=hiveorc')  
format 'custom' (formatter='pxfwritable_import'); 

create external table referrer_dim(
referrer_sk         int8,                 
referrer_txt        text,                
referrer_domain_nm  varchar(200) 
)
location ('pxf://hdp1:51200/test.referrer_dim?profile=hiveorc')  
format 'custom' (formatter='pxfwritable_import'); 

create external table status_code_dim(
status_cd           varchar(100),                            
client_error_flg    int4,                                     
status_cd_desc      text,                                  
server_error_flg    int4
) 
location ('pxf://hdp1:51200/test.status_code_dim?profile=hiveorc')  
format 'custom' (formatter='pxfwritable_import'); 

create external table page_click_fact(
visitor_id                varchar(100),                            
detail_tm                 timestamp,                               
page_click_dt             date,                                    
page_sk                   int8,                                  
client_session_dt         date, 
previous_page_sk          int8,                                  
referrer_sk               int8,                                  
next_page_sk              int8,                                  
status_cd                 varchar(100),                            
browser_sk                int8,                                  
bytes_received_cnt        int8,                                  
bytes_sent_cnt            int8,                                  
client_detail_tm          timestamp,                               
entry_point_flg           int4,                                     
exit_point_flg            int4,                                     
ip_address                varchar(20),                             
query_string_txt          text,                                  
seconds_spent_on_page_cnt int4,                                     
sequence_no               int4,                                     
requested_file_txt        text    
) 
location ('pxf://hdp1:51200/test.page_click_fact?profile=hiveorc')  
format 'custom' (formatter='pxfwritable_import');

说明：HAWQ 外部表使用 PXF 协议，指向相应的 hive 表。

5. 建立 HAWQ 内部表

set search_path=public;

create table date_dim(
cal_dt              date,      
day_in_cal_yr_no    int4,
day_of_week_no      int4,
start_of_month_dt   date, 
start_of_quarter_dt date,     
start_of_week_dt    date,         
start_of_year_dt    date) with (compresstype=snappy,appendonly=true); 

create table browser_dim(
browser_sk              int8,                                  
browser_nm              varchar(100),                            
browser_version_no      varchar(100),                            
flash_version_no        varchar(100),                            
flash_enabled_flg       int,                                     
java_version_no         varchar(100),                            
platform_desc           text,                                  
java_enabled_flg        int,    
java_script_enabled_flg int,    
cookies_enabled_flg     int,                                     
user_language_cd        varchar(100),         
screen_color_depth_no   varchar(100),            
screen_size_txt         text   
) with (compresstype=snappy,appendonly=true); 

create table page_dim(
page_sk             int8,                                  
domain_nm           varchar(200),                            
reachability_cd     text,                           
page_desc           text,                                  
protocol_nm         varchar(20) 
) with (compresstype=snappy,appendonly=true); 

create table referrer_dim(
referrer_sk         int8,                 
referrer_txt        text,                
referrer_domain_nm  varchar(200) 
) with (compresstype=snappy,appendonly=true); 

create table status_code_dim(
status_cd           varchar(100),                            
client_error_flg    int4,                                     
status_cd_desc      text,                                  
server_error_flg    int4
) with (compresstype=snappy,appendonly=true); 

create table page_click_fact(
visitor_id                varchar(100),                            
detail_tm                 timestamp,                               
page_click_dt             date,                                    
page_sk                   int8,                                  
client_session_dt         date, 
previous_page_sk          int8,                                  
referrer_sk               int8,                                  
next_page_sk              int8,                                  
status_cd                 varchar(100),                            
browser_sk                int8,                                  
bytes_received_cnt        int8,                                  
bytes_sent_cnt            int8,                                  
client_detail_tm          timestamp,                               
entry_point_flg           int4,                                     
exit_point_flg            int4,                                     
ip_address                varchar(20),                             
query_string_txt          text,                                  
seconds_spent_on_page_cnt int4,                                     
sequence_no               int4,                                     
requested_file_txt        text    
) with (compresstype=snappy,appendonly=true);

说明：内部表结构定义与 hive 表等价，使用 snappy 压缩的行存储格式。

6. 生成 HAWQ 内部表数据

insert into date_dim select * from hcatalog.test.date_dim;
insert into browser_dim select * from hcatalog.test.browser_dim; 
insert into page_dim select * from hcatalog.test.page_dim; 
insert into referrer_dim select * from hcatalog.test.referrer_dim; 
insert into status_code_dim select * from hcatalog.test.status_code_dim; 
insert into page_click_fact select * from hcatalog.test.page_click_fact;

说明：通过 HCatalog 直接查询 hive 表，插入到 HAWQ 内部表中。snappy 压缩后的各表对应的 HDFS 文件大小如下：

6.2 K   /hawq_data/16385/177422/177677
3.3 M   /hawq_data/16385/177422/177682
23.9 M  /hawq_data/16385/177422/177687
39.3 M  /hawq_data/16385/177422/177707
1.8 K   /hawq_data/16385/177422/177726
7.9 G   /hawq_data/16385/177422/177731

7. 分析 HAWQ 内部表

analyze date_dim;
analyze browser_dim;
analyze page_dim;
analyze referrer_dim;
analyze status_code_dim;
analyze page_click_fact;

六、执行查询

分别在 hive 表、HAWQ 外部表、HAWQ 内部表上执行以下 5 个查询语句，记录执行时间。

1. 查询给定周中 support.sas.com 站点上访问最多的目录

-- hive查询
select top_directory, count(*) as unique_visits     
  from (select distinct visitor_id, substr(requested_file_txt,1,10) top_directory  
          from page_click_fact, page_dim, browser_dim  
         where domain_nm = 'support.sas.com'   
           and flash_enabled_flg=1   
           and weekofyear(detail_tm) = 19   
           and year(detail_tm) = 2017  
       ) directory_summary  
 group by top_directory  
 order by unique_visits;  

-- HAWQ查询，只是用extract函数代替了hive的weekofyear和year函数，与hive的查询语句等价。
select top_directory, count(*) as unique_visits     
  from (select distinct visitor_id, substr(requested_file_txt,1,10) top_directory  
          from page_click_fact, page_dim, browser_dim  
         where domain_nm = 'support.sas.com'   
           and flash_enabled_flg=1   
           and extract(week from detail_tm) = 19   
           and extract(year from detail_tm) = 2017  
       ) directory_summary  
 group by top_directory  
 order by unique_visits;

2. 查询各月从 www.google.com 访问的页面

-- hive查询
select domain_nm, requested_file_txt, count(*) as unique_visitors, month  
  from (select distinct domain_nm, requested_file_txt, visitor_id, month(detail_tm) as month  
          from page_click_fact, page_dim, referrer_dim   
         where domain_nm = 'support.sas.com'   
           and referrer_domain_nm = 'www.google.com'  
       ) visits_pp_ph_summary  
 group by domain_nm, requested_file_txt, month  
 order by domain_nm, requested_file_txt, unique_visitors desc, month asc;  

-- HAWQ查询，只是用extract函数代替了hive的month函数，与hive的查询语句等价。
select domain_nm, requested_file_txt, count(*) as unique_visitors, month  
  from (select distinct domain_nm, requested_file_txt, visitor_id, extract(month from detail_tm) as month  
          from page_click_fact, page_dim, referrer_dim   
         where domain_nm = 'support.sas.com'   
           and referrer_domain_nm = 'www.google.com'  
       ) visits_pp_ph_summary  
 group by domain_nm, requested_file_txt, month  
 order by domain_nm, requested_file_txt, unique_visitors desc, month asc;

3. 给定年份 support.sas.com 站点上的搜索字符串计数

-- hive查询
select query_string_txt, count(*) as count  
  from page_click_fact, page_dim  
 where query_string_txt <> ''   
   and domain_nm='support.sas.com'   
   and year(detail_tm) = '2017'  
 group by query_string_txt  
 order by count desc;
 
-- HAWQ查询，只是用extract函数代替了hive的year函数，与hive的查询语句等价。
select query_string_txt, count(*) as count  
  from page_click_fact, page_dim  
 where query_string_txt <> ''   
   and domain_nm='support.sas.com'   
   and extract(year from detail_tm) = '2017'  
 group by query_string_txt  
 order by count desc;

4. 查询使用 Safari 浏览器访问每个页面的人数

-- hive查询
select domain_nm, requested_file_txt, count(*) as unique_visitors  
  from (select distinct domain_nm, requested_file_txt, visitor_id  
          from page_click_fact, page_dim, browser_dim  
         where domain_nm='support.sas.com'   
           and browser_nm like '%Safari%'   
           and weekofyear(detail_tm) = 19   
           and year(detail_tm) = 2017  
       ) uv_summary  
 group by domain_nm, requested_file_txt  
 order by unique_visitors desc;  
 
-- HAWQ查询，只是用extract函数代替了hive的weekofyear和year函数，与hive的查询语句等价。
select domain_nm, requested_file_txt, count(*) as unique_visitors  
  from (select distinct domain_nm, requested_file_txt, visitor_id  
          from page_click_fact, page_dim, browser_dim  
         where domain_nm='support.sas.com'   
           and browser_nm like '%Safari%'   
           and extract(week from detail_tm) = 19   
           and extract(year from detail_tm) = 2017  
       ) uv_summary  
 group by domain_nm, requested_file_txt  
 order by unique_visitors desc;

5. 查询给定周中 support.sas.com 站点上浏览超过 10 秒的页面

-- hive查询
select domain_nm, requested_file_txt, count(*) as unique_visits  
  from (select distinct domain_nm, requested_file_txt, visitor_id  
          from page_click_fact, page_dim  
         where domain_nm='support.sas.com'   
           and weekofyear(detail_tm) = 19   
           and year(detail_tm) = 2017   
           and seconds_spent_on_page_cnt > 10 
       ) visits_summary  
 group by domain_nm, requested_file_txt  
 order by unique_visits desc;  

-- HAWQ查询，只是用extract函数代替了hive的weekofyear和year函数，与hive的查询语句等价。 
select domain_nm, requested_file_txt, count(*) as unique_visits  
  from (select distinct domain_nm, requested_file_txt, visitor_id  
          from page_click_fact, page_dim  
         where domain_nm='support.sas.com'   
           and extract(week from detail_tm) = 19   
           and extract(year from detail_tm) = 2017   
           and seconds_spent_on_page_cnt > 10 
       ) visits_summary  
 group by domain_nm, requested_file_txt  
 order by unique_visits desc;

七、测试结果

Hive、HAWQ 外部表、HAWQ 内部表查询时间对比如表2 所示，每种查询情况执行三次取平均值。

查询	Hive（秒）	HAWQ外部表（秒）	HAWQ内部表（秒）
1	74.337	304.134	19.232
2	169.521	150.882	3.446
3	73.482	101.216	18.565
4	66.367	359.778	1.217
5	60.341	118.329	2.789

表2

从图2 中的对比可以看到，HAWQ 内部表比 Hive on Tez 快的多（4-50倍）。同样的查询，在 HAWQ 的 Hive 外部表上执行却很慢。因此，在执行分析型查询时最好使用 HAWQ 内部表。如果不可避免地需要使用外部表，为了获得满意的查询性能，需要保证外部表数据量尽可能小。同时要使查询尽可能简单，尽量避免在外部表上执行聚合、分组、排序等复杂操作。

图2