FlinkSql系列1之窗口聚合

FlinkSql之窗口聚合

---

前言

使用flink datastream我们都知道有可以进行开窗聚合处理，今天记录一下flinksql的窗口聚合学习。

一、窗口聚合的分类

滚动窗口(Tumble)
滑动窗口(HOP)
Session窗口(Session)
渐进式窗口(CUMULATE)

二、窗口实际案例与实践

2.1.滚动窗口（TUMBLE）

创建源表

CREATE TABLE source_table(
--维度数据
`city_id` STRING,
--用户id
`user_id` BIGINT,
--金额
`price` BIGINT,
--事件时间戳
`rowtime` AS CAST(CURRENT_TIMESTAMP AS timestamp(3)),
--watermark设置
WATERMARK FOR rowtime AS rowtime - INTERVAL '3' second
)WITH (
 'connector' = 'datagen',
 'rows-per-second' = '5',
 'fields.city_id.length' = '1',
 'fields.user_id.min' = '1',
 'fields.user_id.max' = '50',
 'fields.price.min' = '1',
 'fields.price.max' = '50'
);

创建目标表

CREATE TABLE sink_table(
--维度数据
`city_id` STRING,
--pv pagevisit
`pv` BIGINT,
--总金额
`sum_price` BIGINT,
--最大金额
`max_price` BIGINT,
--最小金额
`min_price` BIGINT,
--uv 去重
`uv` BIGINT,
--窗口开始时间
`winodwstart` BIGINT,
--窗口结束时间
`windowend` BIGINT
) WITH (
'connector'='print'
)

group window aggregation

flink 1.13版本之前支持这种方式

INSERT INTO sink_table
SELECT
city_id,
COUNT(user_id) as pv,
COUNT(price) as sum_price,
MAX(price) as max_price,
MIN(price) as min_price,
COUNT(DISTINCT user_id) as uv,
UNIX_TIMESTAMP(CAST(TUMBLE_START(rowtime,INTERVAL '5' SECOND) AS STRING)) * 1000 as winodwstart,
UNIX_TIMESTAMP(CAST(TUMBLE_END(rowtime,INTERVAL '5' SECOND) AS STRING)) * 1000 as windowend
FROM source_table
GROUP BY
city_id,
TUMBLE(rowtime,INTERVAL '5' SECOND)

window tvf方案（只支持Streaming）

INSERT INTO sink_table
SELECT
city_id,
COUNT(user_id) as pv,
COUNT(DISTINCT user_id) as uv,
COUNT(price) as sum_price,
MAX(price) as max_price,
MIN(price) as min_price,
UNIX_TIMESTAMP(CAST(window_start AS STRING)) * 1000 as window_start,
UNIX_TIMESTAMP(CAST(window_end AS STRING)) * 1000 as window_end
FROM TABLE(TUMBLE(
TABLE source_table,
DESCRIPTOR(rowtime),
INTERVAL '5' SECOND
))
GROUP BY
window_start,
window_end,
city_id

第⼀个参数 TABLE source_table 声明数据源表；
第⼆个参数 DESCRIPTOR(row_time) 声明数据源的时间戳；
第三个参数 INTERVAL ‘5’ SECOND 声明滚动窗⼝⼤⼩为 5s。

2. 2.测试运行

两种方法都可以，但是尽量1.13版本之后建议使用window tvf方案

2.3 滑动窗口（HOP）

创建目标表

CREATE TABLE sink_table1(
--维度数据
`city_id` STRING,
--uv 去重
`uv` BIGINT
--窗口开始时间
`window_start`,
--窗口结束时间
`window_end`
)WITH (
'connector'='print'
)

group window aggregation

INSERT INTO sink_table1
SELECT
city_id,
COUNT(DISTINCT user_id) as uv,
UNIX_TIMESTAMP(CAST(HOP_START(rowtime,INTERVAL '5' SECOND,INTERVAL '10' SECOND) AS STRING)) * 1000 as window_start,
UNIX_TIMESTAMP(CAST(HOP_END(rowtime,INTERVAL '5' SECOND,INTERVAL '10' SECOND) AS STRING)) * 1000 as window_end
FROM source_table
GROUP BY
city_id,
--十秒的滑动窗口，每5秒进行一次滑动，并触发计算，有窗口重叠
HOP(rowtime,INTERVAL '5' SECOND,INTERVAL '10' SECOND);

window tvf方案

INSERT INTO sink_table1
SELECT
city_id,
COUNT(DISTINCT user_id) as uv,
UNIX_TIMESTAMP(CAST(window_start AS STRING)) * 1000 as window_start,
UNIX_TIMESTAMP(CAST(window_end AS STRING)) * 1000 as window_end
--十秒的滑动窗口，每5秒进行一次滑动，并触发计算，有窗口重叠
FROM TABLE(HOP(
TABLE source_table,
DESCRIPTOR(rowtime),
INTERVAL '5' SECOND,
INTERVAL '10' SECOND
))
GROUP BY 
city_id,
window_start,
window_end

2.4 测试结果

2.5 Session窗口（Session）

创建目标表

CREATE TABLE sink_table2(
--维度数据
`city_id` STRING,
--pv 计算每个session窗口内的商品数量
`pv` BIGINT
--窗口开始时间
`window_start`,
--窗口结束时间
`window_end`
)WITH(
'connector'='print'
)

执行sql统计每五秒的会话窗口 windowaggregation

INSERT INTO sink_table2
SELECT
city_id,
COUNT(user_id) as pv,
UNIX_TIMESTAMP(CAST(SESSION_START(rowtime,INTERVAL '5' SECOND) AS STRING)) * 1000 as window_start,
UNIX_TIMESTAMP(CAST(SESSION_END(rowtime,INTERVAL '5' SECOND) AS STRING))* 1000 as window_end
FROM source_table
GROUP BY 
city_id,
SESSION(rowtime,INTERVAL '5' SECOND)

2.6 测试结果

使用Window tvf的方式测试报错

2.7 渐进式窗口（CUMULATE）

创建源表

CREATE TABLE  source_table1(
--用户id
`user_id` STRING,
--消费金额
`money` BIGINT,
--事件时间
`rowtime` AS CAST(CURRENT_TIMESTAMP AS TIMESTAMP(3)),
--指定watermark
WATERMARK FOR rowtime AS rowtime - INTERVAL '2' SECOND
)WITH (
 'connector' = 'datagen',
 'rows-per-second' = '5',
 'fields.user_id.length' = '1',
 'fields.money.min' = '1',
 'fields.money.max' = '50'
)

创建目标表

CREATE TABLE sink_table3(
--用户id总数 去重
`usernum` STRING,
--窗口内总金额数量
`summoney` BIGINT,
--窗口开始时间
`window_start` BIGINT，
--窗口结束时间
`window_end` BIGINT
)WITH (
'connector'='print'
)

执行sql插入，注意这里1.13(streaming)只能使用window tvf模式

INSERT INTO sink_table3
SELECT
COUNT(DISTINCT user_id) as usernum,
COUNT(money) as summoney,
UNIX_TIMESTAMP(CAST(window_start AS STRING)) * 1000 as window_start,
UNIX_TIMESTAMP(CAST(window_end AS STRING)) * 1000 as window_end
FROM TABLE(CUMULATE(
TABLE source_table1,
DESCRIPTOR(rowtime),
INTERVAL ‘5’ SECOND,
INTERVAL ‘20’ SECOND
))
GROUP BY
window_start,
window_end

第⼀个参数 TABLE source_table 声明数据源表；
第⼆个参数 DESCRIPTOR(row_time) 声明数据源的时间戳；
第三个参数 INTERVAL ‘5’ SECOND 声明渐进式窗⼝触发的渐进步⻓为 5s。
第四个参数 INTERVAL ‘20’ SECOND 声明整个渐进式窗⼝的⼤⼩为 20 s，到了20s新开⼀个窗⼝重新累计。

2.8 测试结果

总结

本次对于flinksql里面的四个窗口分别进行了演示
1.滚动窗⼝:将每个元素指定给指定窗⼝⼤⼩的窗⼝。滚动窗⼝具有固定⼤⼩，且不重叠
2.滑动窗口：滑动窗⼝也是将元素指定给固定⻓度的窗⼝。与滚动窗⼝功能⼀样，也有窗⼝⼤⼩的概 念。不⼀样的地⽅在于，滑动窗⼝有另⼀个参数控制窗⼝计算的频率（滑动窗⼝滑动的步⻓）。因此，如果 滑动的步⻓⼩于窗⼝⼤⼩，则滑动窗⼝之间每个窗⼝是可以重叠。在这种情况下，⼀条数据就会分配到多个 窗⼝当中。
3.会话窗口：Session 时间窗⼝和滚动、滑动窗⼝不⼀样，其没有固定的持续时间，如果在定义 的间隔期（Session Gap）内没有新的数据出现，则 Session 就会窗⼝关闭
4.渐进式窗口：渐进式窗⼝在其实就是 固定窗⼝间隔内提前触发的的 滚动窗⼝ ，其实就是 Tumble Window + early-fire 的⼀个事件时间的版本。

这里需要注意尽量使用Window TVF模式，但是会话窗口的好像只能支持Window AGGREGATION.其他的均可以只是Window TVF模式。