Flink的窗口机制

窗口机制

tumble（滚动窗口）
hop（滑动窗口）
session（会话窗口）
cumulate（渐进式窗口）
Over（聚合窗口）

滚动窗口（tumble）

概念

滚动窗口（tumble）：窗口大小 = 滑动距离。

它的窗口是紧密排布的，中间没有任何的数据重复和丢失。

案例 - SQL

--创建表
CREATE TABLE source_table ( 
 user_id STRING, 
 price BIGINT,
 `timestamp` bigint,
 row_time AS TO_TIMESTAMP(FROM_UNIXTIME(`timestamp`)),
 watermark for row_time as row_time - interval '0' second
) WITH (
  'connector' = 'socket',
  'hostname' = 'node1',        
  'port' = '9999',
  'format' = 'csv'
);

--语法
tumble(row_time,时间间隔)，比如，如下的sql
tumble(row_time,interval '5' second)，每隔5秒滚动一次。

--业务查询逻辑（传统方式）
select 
user_id,
count(*) as pv,
sum(price) as sum_price,
UNIX_TIMESTAMP(CAST(tumble_start(row_time, interval '5' second) AS STRING)) * 1000  as window_start,
UNIX_TIMESTAMP(CAST(tumble_end(row_time, interval '5' second) AS STRING)) * 1000  as window_end
from source_table
group by
    user_id,
    tumble(row_time, interval '5' second);

运行结果如下：

案例 - DataStream API

package day04;

import org.apache.flink.api.common.eventtime.SerializableTimestampAssigner;
import org.apache.flink.api.common.eventtime.WatermarkStrategy;
import org.apache.flink.api.common.functions.MapFunction;
import org.apache.flink.api.common.typeinfo.Types;
import org.apache.flink.api.java.tuple.Tuple2;
import org.apache.flink.api.java.tuple.Tuple3;
import org.apache.flink.streaming.api.datastream.DataStreamSource;
import org.apache.flink.streaming.api.datastream.KeyedStream;
import org.apache.flink.streaming.api.datastream.SingleOutputStreamOperator;
import org.apache.flink.streaming.api.datastream.WindowedStream;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.windowing.assigners.TumblingEventTimeWindows;
import org.apache.flink.streaming.api.windowing.time.Time;
import org.apache.flink.streaming.api.windowing.windows.TimeWindow;

/**
 * @desc: 演示基于事件时间的滚动窗口。
 * 从socket数据源中接收数据（id,price,ts）
 */
public class Demo01_TumbleWindow {
    public static void main(String[] args) throws Exception {
        //1.创建流式执行环境
        StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        env.setParallelism(1);
        //2.数据源
        DataStreamSource<String> source = env.socketTextStream("node1", 9999);
        //3.数据处理
        //3.1 数据切分，转成Tuple3
        //3.2 分配窗口
        /**
         * WatermarkStrategy的策略有四种：
         * forMonotonousTimestamps，单调递增水印
         * forBoundedOutOfOrderness，允许乱序数据(数据迟到)的水印
         * forGenerator，自定义水印
         * noWatermarks，没有水印
         */
        SingleOutputStreamOperator<Tuple3<String, Integer, Long>> mapAndWatermarkStream = source.map(new MapFunction<String, Tuple3<String, Integer, Long>>() {
            @Override
            public Tuple3<String, Integer, Long> map(String value) throws Exception {
                String[] lines = value.split(",");
                /**
                 * lines分为3个字段：String id，Integer price，Long ts
                 */
                return Tuple3.of(lines[0], Integer.valueOf(lines[1]), Long.parseLong(lines[2]));
            }
        }).assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple3<String, Integer, Long>>forMonotonousTimestamps()
                .withTimestampAssigner(new SerializableTimestampAssigner<Tuple3<String, Integer, Long>>() {
            @Override
            public long extractTimestamp(Tuple3<String, Integer, Long> element, long recordTimestamp) {
                //这个方法就是用来标识（标记）哪个列是用来表示时间戳的
                return element.f2 * 1000L;
            }
        }));

        mapAndWatermarkStream.print("源数据：");

        //3.3 根据id进行分组
        KeyedStream<Tuple3<String, Integer, Long>, String> keyedStream = mapAndWatermarkStream.keyBy(value -> value.f0);
        //3.4分组后，进行窗口划分
        //3.5 划分窗口后，对窗口内的数据进行sum
        //3.6为了显示的友好性，我们对Tuple3<String,Integer,Long>进行转换成Tuple2<String,Integer>
        SingleOutputStreamOperator<Tuple2<String, Integer>> result = keyedStream.window(TumblingEventTimeWindows.of(Time.seconds(5)))
                .sum(1)
                .map(value -> Tuple2.of(value.f0, value.f1)).returns(Types.TUPLE(Types.STRING, Types.INT));

        //4.数据输出
        result.printToErr("聚合后的数据：");
        //5.启动流式任务
        env.execute();
    }
}

运行结果如下：

SQL案例TVF写法

--语法,跟3个参数：
--参数1：表名
--参数2：表中事件时间列
--参数3：窗口大小
from table(tumble(table source,descriptor(row_time),interval '5' second))

--业务逻辑
SELECT 
    user_id,
    UNIX_TIMESTAMP(CAST(window_start AS STRING)) * 1000 as window_start,
    UNIX_TIMESTAMP(CAST(window_end AS STRING)) * 1000 as window_end,
    sum(price) as sum_price
FROM TABLE(TUMBLE(
        TABLE source_table
        , DESCRIPTOR(row_time)
        , INTERVAL '5' SECOND))
GROUP BY window_start, 
      window_end,
      user_id;
--window_start，window_end是新写法的关键字

      
--对照前面的方式：
select 
user_id,
count(*) as pv,
sum(price) as sum_price,
UNIX_TIMESTAMP(CAST(tumble_start(row_time, interval '5' second) AS STRING)) * 1000  as window_start,
UNIX_TIMESTAMP(CAST(tumble_end(row_time, interval '5' second) AS STRING)) * 1000  as window_end
from source_table
group by
    user_id,
    tumble(row_time, interval '5' second);

运行结果如下：

滑动窗口（hop）

概念

滑动窗口：滑动距离不等于窗口大小。

（1）如果滑动距离小于窗口大小，则会产生数据重复

（2）如果滑动距离等于窗口大小，这就是滚动窗口

（3）如果滑动距离大于窗口大小，则会产生数据丢失（不考虑）

案例 - SQL

--创建表
CREATE TABLE source_table ( 
 user_id STRING, 
 price BIGINT,
 `timestamp` bigint,
 row_time AS TO_TIMESTAMP(FROM_UNIXTIME(`timestamp`)),
 watermark for row_time as row_time - interval '0' second
) WITH (
  'connector' = 'socket',
  'hostname' = 'node1',        
  'port' = '9999',
  'format' = 'csv'
);

--语法
hop(事件时间列，滑动间隔，窗口大小)
hop(row_time,interval '2' second, interval '5' second)

--业务SQL
SELECT user_id,
UNIX_TIMESTAMP(CAST(hop_start(row_time, interval '2' SECOND, interval '5' SECOND) AS STRING)) * 1000 as window_start,
UNIX_TIMESTAMP(CAST(hop_end(row_time, interval '2' SECOND, interval '5' SECOND) AS STRING)) * 1000 as window_end, 
    sum(price) as sum_price
FROM source_table
GROUP BY user_id
    , hop(row_time, interval '2' SECOND, interval '5' SECOND);--每隔两秒滑动一次

运行结果如下：

SQL案例TVF写法

--语法
--table:表名
--descriptor：事件时间列
--滑动距离：interval 2 second
--窗口大小：interval 5 second
from table(hop(table 表名,descriptor(事件时间列),滑动间隔，窗口大小))
from table(hop(table source,descriptor(row_time),interval '2' second,interval '5' second))

--业务SQL
SELECT 
    user_id,
UNIX_TIMESTAMP(CAST(window_start AS STRING)) * 1000 as window_start,  
UNIX_TIMESTAMP(CAST(window_end AS STRING)) * 1000 as window_end, 
    sum(price) as sum_price
FROM TABLE(HOP(
        TABLE source_table
        , DESCRIPTOR(row_time)
        , interval '2' SECOND, interval '6' SECOND))
GROUP BY window_start, 
      window_end,
      user_id;

运行结果如下：

案例 - DataStream API

package day04;

import org.apache.flink.api.common.eventtime.SerializableTimestampAssigner;
import org.apache.flink.api.common.eventtime.WatermarkStrategy;
import org.apache.flink.api.common.typeinfo.Types;
import org.apache.flink.api.java.tuple.Tuple2;
import org.apache.flink.api.java.tuple.Tuple3;
import org.apache.flink.streaming.api.datastream.DataStreamSource;
import org.apache.flink.streaming.api.datastream.SingleOutputStreamOperator;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.windowing.assigners.SlidingEventTimeWindows;
import org.apache.flink.streaming.api.windowing.time.Time;

/**
 * @desc: 演示基于事件时间的滑动窗口。
 * 从socket数据源中接收数据（id,price,ts）
 */
public class Demo02_SlideWindow {
    public static void main(String[] args) throws Exception {
        //1.创建流式执行环境
        StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        env.setParallelism(1);
        //2.数据源
        DataStreamSource<String> source = env.socketTextStream("node1", 9999);
        //3.数据处理
        //3.1数据map转换操作，转成Tuple3
        SingleOutputStreamOperator<Tuple3<String, Integer, Long>> mapStream = source.map(value -> {
            /**
             * 由一行数据中，用逗号进行切分【id，price，ts】
             * String id
             * Integer price
             * Long ts
             */
            String[] lines = value.split(",");
            return Tuple3.of(lines[0], Integer.valueOf(lines[1]), Long.parseLong(lines[2]));
        }).returns(Types.TUPLE(Types.STRING, Types.INT, Types.LONG));
        //3.2把Tuple3的数据添加Watermark（monotonousTimestamps）
        /**
         * WatermarkStrategy生成水印的策略有四种：
         * forMonotonousTimestamps，单调递增水印（用的次多）
         * forBoundedOutOfOrderness，运行数据迟到（乱序）（用的最多）
         * forGeneric，自定义水印（不用）
         * noWatermark，没有水印（不用）
         */
        SingleOutputStreamOperator<Tuple3<String, Integer, Long>> watermarks = mapStream.assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple3<String, Integer, Long>>forMonotonousTimestamps()
                .withTimestampAssigner(new SerializableTimestampAssigner<Tuple3<String, Integer, Long>>() {
                    @Override
                    public long extractTimestamp(Tuple3<String, Integer, Long> element, long recordTimestamp) {
                        return element.f2 * 1000L;
                    }
                }));
        //3.3把数据根据id进行分组
        watermarks.print("源数据：");
        SingleOutputStreamOperator<Tuple3<String, Integer, Long>> sumStream = watermarks.keyBy(value -> value.f0)
                //3.4分组之后，设置滑动事件时间窗口，并且制定窗口大小为5秒钟，滑动间隔为2秒。
                /**
                 * SlidingEventTimeWindows，滑动事件时间窗口，带2个参数：
                 * 参数1：窗口大小
                 * 参数2：滑动间隔
                 */
                .window(SlidingEventTimeWindows.of(Time.seconds(5), Time.seconds(2)))
                //3.5对窗口内的数据进行sum操作
                .sum(1);
        //3.6把Tuple3转成了Tuple2（取id和sum的值）
        SingleOutputStreamOperator<Tuple2<String, Integer>> result = sumStream.map(value -> Tuple2.of(value.f0, value.f1))
                .returns(Types.TUPLE(Types.STRING,Types.INT));
        //4.数据输出
        result.printToErr("聚合后的数据：");
        //5.启动流式任务
        env.execute();
    }
}

运行结果如下：

会话窗口（session）

概念

会话窗口：在一个会话周期内，窗口的数据会累积，超过会话周期就会触发窗口的计算，同时开辟下一个新窗口。

注意：

数据本身的事件时间大于窗口间隔，才会触发当前窗口的计算。

案例 - SQL

--创建表
CREATE TABLE source_table ( 
 user_id STRING, 
 price BIGINT,
 `timestamp` bigint,
 row_time AS TO_TIMESTAMP(FROM_UNIXTIME(`timestamp`)),
 watermark for row_time as row_time - interval '0' second
) WITH (
  'connector' = 'socket',
  'hostname' = 'node1',        
  'port' = '9999',
  'format' = 'csv'
);

--语法
session(事件时间列，窗口间隔)
session(row_time,interval '5' second)

--业务SQL
SELECT 
    user_id,
UNIX_TIMESTAMP(CAST(session_start(row_time, interval '5' SECOND) AS STRING)) * 1000 as window_start,
UNIX_TIMESTAMP(CAST(session_end(row_time, interval '5' SECOND) AS STRING)) * 1000 as window_end, 
    sum(price) as sum_price
FROM source_table
GROUP BY user_id
      , session(row_time, interval '5' SECOND);

运行结果如下：

案例 - DataStream API

package day04;

import org.apache.flink.api.common.eventtime.SerializableTimestampAssigner;
import org.apache.flink.api.common.eventtime.WatermarkStrategy;
import org.apache.flink.api.common.functions.MapFunction;
import org.apache.flink.api.common.typeinfo.Types;
import org.apache.flink.api.java.tuple.Tuple2;
import org.apache.flink.api.java.tuple.Tuple3;
import org.apache.flink.streaming.api.datastream.DataStreamSource;
import org.apache.flink.streaming.api.datastream.SingleOutputStreamOperator;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.windowing.assigners.EventTimeSessionWindows;
import org.apache.flink.streaming.api.windowing.time.Time;

/**
 * @desc: 演示基于事件时间的会话窗口。
 * 从socket数据源中接收数据（id,price,ts）
 */
public class Demo03_SessionWindow {
    public static void main(String[] args) throws Exception {
        //1.创建流式执行环境
        StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        env.setParallelism(1);
        //2.数据源
        DataStreamSource<String> source = env.socketTextStream("node1", 9999);
        //3.数据处理
        //3.1数据map转换操作，转成Tuple3
        SingleOutputStreamOperator<Tuple3<String, Integer, Long>> mapStream = source.map(new MapFunction<String, Tuple3<String, Integer, Long>>() {
            @Override
            public Tuple3<String, Integer, Long> map(String value) throws Exception {
                /**
                 * String id
                 * Integer price
                 * Long ts
                 */
                String[] lines = value.split(",");
                return Tuple3.of(lines[0], Integer.valueOf(lines[1]), Long.parseLong(lines[2]));
            }
        });
        //3.2把Tuple3的数据添加Watermark（monotonousTimestamps）
        SingleOutputStreamOperator<Tuple3<String, Integer, Long>> watermarks = mapStream
                .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple3<String, Integer, Long>>forMonotonousTimestamps()
                .withTimestampAssigner(new SerializableTimestampAssigner<Tuple3<String, Integer, Long>>() {
            @Override
            public long extractTimestamp(Tuple3<String, Integer, Long> element, long recordTimestamp) {
                return element.f2 * 1000L;
            }
        }));
        watermarks.print("源数据：");
        //3.3把数据根据id进行分组
        //3.4分组之后，设置会话事件时间窗口，并且指定窗口间隔为5秒钟。
        //3.5对窗口内的数据进行sum操作
        SingleOutputStreamOperator<Tuple3<String, Integer, Long>> sumStream = watermarks.keyBy(value -> value.f0)
                .window(EventTimeSessionWindows.withGap(Time.seconds(5)))
                .sum(1);
        //3.6把Tuple3转成了Tuple2（取id和sum的值）
        SingleOutputStreamOperator<Tuple2<String, Integer>> result = sumStream.map(value -> Tuple2.of(value.f0, value.f1)).returns(Types.TUPLE(Types.STRING, Types.INT));
        //4.数据输出
        result.printToErr("聚合后的结果：");
        //5.启动流式任务
        env.execute();
    }
}

运行结果如下：

渐进式窗口（cumulate）

概念

渐进式窗口：在固定时间内，结果是单调递增的。比如周期性累计求某些指标。

如下图所示：

案例 -TVF方式

--语法
--1.source：表名
--2.事件时间列
--3.时间间隔，每隔多久把窗口内的数据统计一次
--4.窗口大小，窗口的时间长度
from table(cumulate(table source,descriptor(事件时间列),时间间隔，窗口大小))

--建表
CREATE TABLE source_table (
    -- 用户 id
    user_id BIGINT,
    -- 用户
    money BIGINT,
    -- 事件时间戳
    row_time AS cast(CURRENT_TIMESTAMP as timestamp(3)),
    -- watermark 设置
    WATERMARK FOR row_time AS row_time - INTERVAL '0' SECOND
) WITH (
  'connector' = 'datagen',
  'rows-per-second' = '10',
  'fields.user_id.min' = '1',
  'fields.user_id.max' = '100000',
  'fields.money.min' = '1',
  'fields.money.max' = '100000'
);

--业务SQL
SELECT 
FROM_UNIXTIME(UNIX_TIMESTAMP(CAST(window_start AS STRING)))  as window_start,
    FROM_UNIXTIME(UNIX_TIMESTAMP(CAST(window_end AS STRING))) as window_end, 
    sum(money) as sum_money,
    count(distinct user_id) as count_distinct_id
FROM TABLE(CUMULATE(
       TABLE source_table
       , DESCRIPTOR(row_time)
       , INTERVAL '5' SECOND
       , INTERVAL '30' SECOND))
GROUP BY
    window_start, 
    window_end;

--上面的cumulate窗口翻译如下：
--每隔5秒，统计最近30秒内的数据（总金额，用户量）

运行结果如下：

补充：

1.渐进式窗口没有普通SQL的写法。

2.渐进式窗口没有DataStream API。

Over 窗口

聚合窗口，用的不多。

Over 窗口分为两类：

时间区间范围
行数

时间区间范围

--语法
--range：范围，
--between ... and ... ： 在...之间
--INTERVAL '1' HOUR PRECEDING：一小时前
--CURRENT ROW：当前行
RANGE BETWEEN 起始时间范围 AND CURRENT ROW
RANGE BETWEEN INTERVAL '1' HOUR PRECEDING AND CURRENT ROW

--建表SQL
CREATE TABLE source_table (
    order_id BIGINT,
    product BIGINT,
    amount BIGINT,
    order_time as cast(CURRENT_TIMESTAMP as TIMESTAMP(3)),
    WATERMARK FOR order_time AS order_time - INTERVAL '0' SECOND
) WITH (
  'connector' = 'datagen',
  'rows-per-second' = '1',
  'fields.order_id.min' = '1',
  'fields.order_id.max' = '2',
  'fields.amount.min' = '1',
  'fields.amount.max' = '10',
  'fields.product.min' = '1',
  'fields.product.max' = '2'
);

--业务SQL
SELECT product, order_time, amount,
  SUM(amount) OVER (
    PARTITION BY product
    ORDER BY order_time
    -- 标识统计范围是一个 product 的最近 30秒的数据
    RANGE BETWEEN INTERVAL '1' HOUR PRECEDING AND CURRENT ROW
  ) AS one_hour_prod_amount_sum
FROM source_table;

运行结果如下：

行数

--语法
--ROWS:行数
--行数 preceding：往前多少行
--：CURRENT ROW：当前行
ROWS BETWEEN 行数 PRECEDING AND CURRENT ROW
--统计最近100行到当前行的指标
ROWS BETWEEN 100 PRECEDING AND CURRENT ROW
对比前面的时间区间语法：
RANGE BETWEEN INTERVAL '1' HOUR PRECEDING AND CURRENT ROW

--创建表
CREATE TABLE source_table (
    order_id BIGINT,
    product BIGINT,
    amount BIGINT,
    order_time as cast(CURRENT_TIMESTAMP as TIMESTAMP(3)),
    WATERMARK FOR order_time AS order_time - INTERVAL '0' SECOND
) WITH (
  'connector' = 'datagen',
  'rows-per-second' = '1',
  'fields.order_id.min' = '1',
  'fields.order_id.max' = '2',
  'fields.amount.min' = '1',
  'fields.amount.max' = '2',
  'fields.product.min' = '1',
  'fields.product.max' = '2'
);

--业务SQL
SELECT product, order_time, amount,
  SUM(amount) OVER (
    PARTITION BY product
    ORDER BY order_time
    -- 标识统计范围是一个 product 的最近 5 行数据
    ROWS BETWEEN 100 PRECEDING AND CURRENT ROW
  ) AS one_hour_prod_amount_sum
FROM source_table;

运行结果如下：

补充：

1.Over窗口没有传统窗口的SQL写法。

2.Over窗口没有DataStream API的写法。

标签： flink 大数据

本文转载自: https://blog.csdn.net/sijiwang95/article/details/130060606
版权归原作者 javaisGod_s 所有，如有侵权，请联系我们删除。

Flink的窗口机制

窗口机制

滚动窗口（tumble）

滑动窗口（hop）

会话窗口（session）

渐进式窗口（cumulate）

Over 窗口

时间区间范围

行数

发表评论

“Flink的窗口机制”的评论:

关于作者

overfit同步小助手

相关阅读

文章导航