4、聚合操作
4.1、group by 操作
group by操作是实际业务场景(如实时报表、实时大屏等)中使用最为频繁的操作。通常实时聚合的主要源头数据流不会包含丰富的上下文信息,而是经常需要实时关联相关
相关的维度表,并针对这些扩展的、丰富维度属性进行各种业务的统计。
在下面的实例中,订单流通过买家id关联了买家维度表,获取其所在省份信息,然后实时统计每天各个省份的iPhone销量信息。
---从源头接收订单实时流
create table test_order_stream (
gmt_create varchar,
gmt_modified varchar,
order_id bigint,
buyer_id bigint,
seller_id bigint,
item_id bigint,
json_object varchar,
order_type varchar,
category_name varchar,
sub_category_name varchar
) with (
type = 'datahub',
endpoint = 'http://dh-et2.aliyun-inc.com',
project = ' your_project',
topic = 'test_topic_1',
accessId = 'your_access_id',
accessKey = 'your_access_key',
startTime = '2018-08-08-00:00:00'
);
----定义rds买家维度表
create table rds_dim_buyer(
buyer_id int,
age int,
province varchar,
star_level varchar,
primary key(buyer_id),
period for system_time ---定义了维度表的变化周期,即是一张变化的表
) with (
type = 'rds',
url = 'your_mySQL_url',
tableName = 'your_table_name',
userName = 'your_user_name',
password = 'your_password'
);
---订单流关联买家维度表获取买家所在省份,并过滤非iPhone订单
create view tmp_order as
select ord.order_id,
ord.gmt_create as order_create_time,
ord.buyer_id,
byr.age,
byr.provice,
byr.star_level
from test_order_stream as ord
left join rds_dim_buyer for system_time s of proctime() as byr
--实际项目中,可能为了避免join热点,对买家维度表做了md5处理,那么join的时候也要做对应处理,
--如,新的join条件可能会变为:
--on concat(substr(md5(ord.buyer_id), 1, 4), '_', ord.order_id) = byr.md5_byr_id
on ord.buyer_id = byr.buyer_id
where ord.category_name = '手机'
and ord.sub_category_name='iPhone';
---定义rds的结果表
create table rds_mobile_orders(
order_create_day varchar,
province varchar,
iphone_order_count int,
primary key(order_create_day,province)
) with (
type = 'rds',
url = 'your_mysql_url',
tableName = ' your_table_name',
userName = 'your_user_name',
password = 'your_password'
);
---安装天、省份汇总每天iphone手机销量
inert into rds_mobile_orders
select
substring(order_create_time, 1, 10) as order_create_day
province,
count(distinct order_id) as iphone_order_count
from tem_order
group by substring(order_create_time, 1, 10) ,province;
4.2、窗口 操作
group by操作的是全局窗口,阿里云Stream SQL还支持包含滑动、滚动、session等的窗口操作,下面以event time的滑动窗口为例介绍窗口操作。
针对event time操作必须首先定义watermark,直接在订单源头流定义即可,hop(datetime, slide, size)函数定义滑动窗口,其中datetime为时间列,slide为滑动间隔,size为窗口大小,
HOP_START则获取到窗口的开始时间,对上述的group by操作进行改动的实例如下,其业务含义为为每一小时统计过去24小时每个省份的的iphone手机销量。
---从源头接收订单实时流
create table test_order_stream (
gmt_create varchar,
gmt_modified varchar,
order_id bigint,
buyer_id bigint,
seller_id bigint,
item_id bigint,
json_object varchar,
order_type varchar,
category_name varchar,
sub_category_name varchar,
WATERMARK mywatermark FOR gmt_modified as withOffset(gmt_modified,1000)
) with (
type = 'datahub',
endpoint = 'http://dh-et2.aliyun-inc.com',
project = ' your_project',
topic = 'test_topic_1',
accessId = 'your_access_id',
accessKey = 'your_access_key',
startTime = '2018-08-08-00:00:00'
);
----定义rds买家维度表
create table rds_dim_buyer(
buyer_id int,
age int,
province varchar,
star_level varchar,
primary key(buyer_id),
period for system_time ---定义了维度表的变化周期,即是一张变化的表
) with (
type = 'rds',
url = 'your_mySQL_url',
tableName = 'your_table_name',
userName = 'your_user_name',
password = 'your_password'
);
---订单流关联买家维度表获取买家所在省份,并过滤非iPhone订单
create view tmp_order as
select ord.order_id,
ord.gmt_create as order_create_time,
ord.buyer_id,
byr.age,
byr.provice,
byr.star_level
from test_order_stream as ord
left join rds_dim_buyer for system_time s of proctime() as byr
--实际项目中,可能为了避免join热点,对买家维度表做了md5处理,那么join的时候也要做对应处理,
--如,新的join条件可能会变为:
--on concat(substr(md5(ord.buyer_id), 1, 4), '_', ord.order_id) = byr.md5_byr_id
on ord.buyer_id = byr.buyer_id
where ord.category_name = '手机'
and ord.sub_category_name='iPhone';
---定义rds的结果表
create table rds_mobile_orders(
stat_begin_time varchar,
province varchar,
iphone_order_count int,
primary key(stat_begin_time,province)
) with (
type = 'rds',
url = 'your_mysql_url',
tableName = ' your_table_name',
userName = 'your_user_name',
password = 'your_password'
);
---每一小时统计过去24小时每个省份iPhone手机销量
insert into rds_moble_orders
select
cast(HOP_START(order_modified_time interval '1' hour, interva '1' day)) as TIMESTAMP) as stat_begin_time,
province,
count(distinct order_id) as ihpone_order_count
from tmp_order
group by HOP_START(order_modified_time interval '1' hour, interva '1' day),province;
5、撤回机制
在某些业务场景下,必须考虑撤回,否则计算结果不准确,比如用户排队咨询的场景,如果某用户A从队列1转移到队列2,现在要统计每个队列最终承担的用户咨询量,那么
不考虑撤回将会导致重复计算。
阿里云Stream SQL支持撤回的处理,具体实例如下,其业务含义为统计每个队列最终承担的用户咨询量。
---从源头接收咨询session粒度的实时流
create table test_queue_stream(
gmt_create varchar,
gmt_modified varchar,
session_id bigint,
queue_id bigint,
session_user_id bigint,
session_user_name bigint
) with (
type = 'datahub',
endpoint = 'http://dh-et2.aliyun-inc.com',
project = ' your_project',
topic = 'test_topic_1',
accessId = 'your_access_id',
accessKey = 'your_access_key',
startTime = '2018-08-08-00:00:00'
);
----创建临时表,取每个session的最后一个queue_id,与下面的group by操作一起支持撤回
create view tmp_queue_stream as
select
session_id,
StringLast(queue_id)
from test_queue_stream
group by session_id;
-----定义rds的结果表
create table rds_queue_result(
queue_id varchar,
session_count int,
primary key(queue_id)
) with (
type = 'rds',
url = 'your_mysql_url',
tableName = ' your_table_name',
userName = 'your_user_name',
password = 'your_password'
)
---统计每个队列的排队量,如果用户有队列变更,group by时会撤回,不会重复统计
insert into rds_queue_result
select queue_id,
count(distinct session_id) as session_count
from tmp_queue_stream
group by queue_id;
参考资料:《离线和实时大数据开发实战》