案例用到的测试数据请参考文章:
Flink自定义Source模拟数据流
原文链接:https://blog.csdn.net/m0_52606060/article/details/135436048
概述
在实际应用中,我们经常会遇到来源不同的多条流,需要将它们的数据进行联合处理。所以Flink中合流的操作会更加普遍,对应的API也更加丰富。
联合(Union)
最简单的合流操作,就是直接将多条流合在一起,叫作流的“联合”(union)类似于SQL里面的union。联合操作要求必须流中的数据类型必须相同,合并之后的新流会包括所有流中的元素,数据类型不变。
在代码中,我们只要基于DataStream直接调用.union()方法,传入其他DataStream作为参数,就可以实现流的联合了;得到的依然是一个DataStream:
stream1.union(stream2, stream3, ...)
案例:合并两个订单数据流
package com.zxl.flink;
import com.zxl.bean.Orders;
import com.zxl.datas.OrdersData;
import org.apache.flink.streaming.api.datastream.DataStream;
import org.apache.flink.streaming.api.datastream.DataStreamSource;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
public class DemoTest {
public static void main(String[] args) throws Exception {
//创建Flink流处理执行环境
StreamExecutionEnvironment environment = StreamExecutionEnvironment.getExecutionEnvironment();
//设置并行度为1
environment.setParallelism(1);
//调用Flink自定义Source
// TODO: 2024/1/6 订单数据
DataStreamSource<Orders> ordersDataStreamSource = environment.addSource(new OrdersData());
// TODO: 2024/1/7 为了效果更佳明显先分割成两条数据流
DataStream<Orders> operator01 = ordersDataStreamSource.filter(orders -> orders.getOrder_amount() > 50);
DataStream<Orders> operator02 = ordersDataStreamSource.filter(orders -> orders.getOrder_amount() < 50);
// TODO: 2024/1/7 合流操作
DataStream<Orders> stream = operator01.union(operator02);
stream.print();
environment.execute();
}
}
连接(Connect)
流的联合虽然简单,不过受限于数据类型不能改变,灵活性大打折扣,所以实际应用较少出现。除了联合(union),Flink还提供了另外一种方便的合流操作——连接(connect),这个流的最终效果和SQLunion类似,只不过可以对每个流和输出的流数据进行处理更加灵活。
代码实现:需要分为两步:首先基于一条DataStream调用.connect()方法,传入另外一条DataStream作为参数,将两条流连接起来,得到一个ConnectedStreams;然后再调用同处理方法得到DataStream。这里可以的调用的同处理方法有.map()/.flatMap(),以及.process()方法
package com.zxl.flink;
import com.zxl.bean.Orders;
import com.zxl.bean.Payments;
import com.zxl.datas.OrdersData;
import com.zxl.datas.PaymentData;
import org.apache.flink.streaming.api.datastream.ConnectedStreams;
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.functions.co.CoFlatMapFunction;
import org.apache.flink.streaming.api.functions.co.CoMapFunction;
import org.apache.flink.streaming.api.functions.co.CoProcessFunction;
import org.apache.flink.util.Collector;
public class DemoTest {
public static void main(String[] args) throws Exception {
//创建Flink流处理执行环境
StreamExecutionEnvironment environment = StreamExecutionEnvironment.getExecutionEnvironment();
//设置并行度为1
environment.setParallelism(1);
//调用Flink自定义Source
// TODO: 2024/1/6 订单数据
DataStreamSource<Orders> ordersDataStreamSource = environment.addSource(new OrdersData());
// TODO: 2024/1/6 支付数据
DataStreamSource<Payments> paymentsDataStreamSource = environment.addSource(new PaymentData());
// TODO: 2024/1/7 连接两条流
ConnectedStreams<Orders, Payments> connectedStreams = ordersDataStreamSource.connect(paymentsDataStreamSource);
// TODO: 2024/1/7 对连接的两条流和输出的流进行处理
// TODO: 2024/1/7 通过Map进行处理
SingleOutputStreamOperator<String> mapStream = connectedStreams.map(new CoMapFunction<Orders, Payments, String>() {
@Override
public String map1(Orders orders) throws Exception {
return orders.toString();
}
@Override
public String map2(Payments payments) throws Exception {
return payments.toString();
}
});
// TODO: 2024/1/7 通过flatMap进行处理
SingleOutputStreamOperator<String> flatMaStream = connectedStreams.flatMap(new CoFlatMapFunction<Orders, Payments, String>() {
@Override
public void flatMap1(Orders orders, Collector<String> collector) throws Exception {
collector.collect(orders.toString());
}
@Override
public void flatMap2(Payments payments, Collector<String> collector) throws Exception {
collector.collect(payments.toString());
}
});
// TODO: 2024/1/7 收集订单金额和支付金额大于50元的数据
SingleOutputStreamOperator<String> processStream = connectedStreams.process(new CoProcessFunction<Orders, Payments, String>() {
@Override
public void processElement1(Orders orders, CoProcessFunction<Orders, Payments, String>.Context context, Collector<String> collector) throws Exception {
if (orders.getOrder_amount() > 50) {
collector.collect(orders.toString());
}
}
@Override
public void processElement2(Payments payments, CoProcessFunction<Orders, Payments, String>.Context context, Collector<String> collector) throws Exception {
if (payments.getPayment_amount() > 50) {
collector.collect(payments.toString());
}
}
});
flatMaStream.print();
environment.execute();
}
}
Map进行处理结果
flatMap进行处理结果
process进行处理的结果
窗口联结(Window Join)
Flink为基于一段时间的双流合并专门提供了一个窗口联结算子,可以定义时间窗口,并将两条流中共享一个公共键(key)的数据放在窗口中进行配对处理。
1)窗口联结的调用
窗口联结在代码中的实现,首先需要调用DataStream的.join()方法来合并两条流,得到一个JoinedStreams;接着通过.where()和.equalTo()方法指定两条流中联结的key;然后通过.window()开窗口,并调用.apply()传入联结窗口函数进行处理计算。通用调用形式如下:
stream1.join(stream2)
.where(<KeySelector>)
.equalTo(<KeySelector>)
.window(<WindowAssigner>)
.apply(<JoinFunction>)
上面代码中.where()的参数是键选择器(KeySelector),用来指定第一条流中的key;而.equalTo()传入的KeySelector则指定了第二条流中的key。两者相同的元素,如果在同一窗口中,就可以匹配起来,并通过一个“联结函数”(JoinFunction)进行处理了。这里.window()传入的就是窗口分配器,之前讲到的三种时间窗口都可以用在这里:滚动窗(tumbling window)、滑动窗口(sliding window)和会话窗口(session window)。而后面调用.apply()可以看作实现了一个特殊的窗口函数。注意这里只能调用.apply(),没有其他替代的方法。传入的JoinFunction也是一个函数类接口,使用时需要实现内部的.join()方法。这个方法有两个参数,分别表示两条流中成对匹配的数据。其实仔细观察可以发现,窗口join的调用语法和我们熟悉的SQL中表的join非常相似:SELECT * FROM table1 t1, table2 t2 WHERE t1.id = t2.id; 这句SQL中where子句的表达,等价于inner join … on,所以本身表示的是两张表基于id的“内连接”(inner join)。而Flink中的window join,同样类似于inner join。也就是说,最后处理输出的,只有两条流中数据按key配对成功的那些;如果某个窗口中一条流的数据没有任何另一条流的数据匹配,那么就不会调用JoinFunction的.join()方法,也就没有任何输出了。
package com.zxl.flink;
import com.zxl.bean.Orders;
import com.zxl.bean.Payments;
import com.zxl.datas.OrdersData;
import com.zxl.datas.PaymentData;
import org.apache.flink.api.common.eventtime.SerializableTimestampAssigner;
import org.apache.flink.api.common.eventtime.WatermarkStrategy;
import org.apache.flink.api.common.functions.JoinFunction;
import org.apache.flink.streaming.api.TimeCharacteristic;
import org.apache.flink.streaming.api.datastream.DataStream;
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.TumblingEventTimeWindows;
import org.apache.flink.streaming.api.windowing.time.Time;
import scala.Tuple9;
public class DemoTest {
public static void main(String[] args) throws Exception {
//创建Flink流处理执行环境
StreamExecutionEnvironment environment = StreamExecutionEnvironment.getExecutionEnvironment();
//设置并行度为1
environment.setParallelism(1);
// TODO: 2024/1/7 定义时间语义
environment.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);
//调用Flink自定义Source
// TODO: 2024/1/6 订单数据
DataStreamSource<Orders> ordersDataStreamSource = environment.addSource(new OrdersData());
// TODO: 2024/1/6 支付数据
DataStreamSource<Payments> paymentsDataStreamSource = environment.addSource(new PaymentData());
// TODO: 2024/1/7 配置订单数据水位线
SingleOutputStreamOperator<Orders> ordersWater = ordersDataStreamSource.assignTimestampsAndWatermarks(WatermarkStrategy
// TODO: 2024/1/7 指定watermark生成:升序的watermark,没有等待时间
.<Orders>forMonotonousTimestamps()
.withTimestampAssigner(new SerializableTimestampAssigner<Orders>() {
@Override
public long extractTimestamp(Orders orders, long l) {
return orders.getOrder_date();
}
})
);
// TODO: 2024/1/7 配置支付数据水位线
SingleOutputStreamOperator<Payments> paymentsWater = paymentsDataStreamSource.assignTimestampsAndWatermarks(WatermarkStrategy
.<Payments>forMonotonousTimestamps()
.withTimestampAssigner(new SerializableTimestampAssigner<Payments>() {
@Override
public long extractTimestamp(Payments payments, long l) {
return payments.getPayment_date();
}
})
);
// TODO window join
// 1. 落在同一个时间窗口范围内才能匹配
// 2. 根据keyby的key,来进行匹配关联
// 3. 只能拿到匹配上的数据,类似有固定时间范围的inner join
DataStream<Tuple9> dataStream = ordersWater.join(paymentsWater)
.where(orders -> orders.getOrder_id())
.equalTo(payments -> payments.getOrder_id())
.window(TumblingEventTimeWindows.of(Time.seconds(10)))
.apply(new JoinFunction<Orders, Payments, Tuple9>() {
@Override
public Tuple9 join(Orders orders, Payments payments) throws Exception {
Tuple9<Long, Long, Long, Integer, Integer, Long, Integer, Long, String> tuple9 = new Tuple9<>(orders.getOrder_id(), orders.getUser_id(), orders.getOrder_date(), orders.getOrder_amount(), orders.getProduct_id(), orders.getOrder_num(), payments.getPayment_amount(), payments.getPayment_date(), payments.getPayment_type());
return tuple9;
}
});
dataStream.print("join后的数据");
environment.execute();
}
}
间隔联结(Interval Join)
在有些场景下,我们要处理的时间间隔可能并不是固定的。这时显然不应该用滚动窗口或滑动窗口来处理——因为匹配的两个数据有可能刚好“卡在”窗口边缘两侧,于是窗口内就都没有匹配了;会话窗口虽然时间不固定,但也明显不适合这个场景。基于时间的窗口联结已经无能为力了。
为了应对这样的需求,Flink提供了一种叫作“间隔联结”(interval join)的合流操作。顾名思义,间隔联结的思路就是针对一条流的每个数据,开辟出其时间戳前后的一段时间间隔,看这期间是否有来自另一条流的数据匹配。
间隔联结的原理
间隔联结具体的定义方式是,我们给定两个时间点,分别叫作间隔的“上界”(upperBound)和“下界”(lowerBound);于是对于一条流(不妨叫作A)中的任意一个数据元素a,就可以开辟一段时间间隔:[a.timestamp + lowerBound, a.timestamp + upperBound],即以a的时间戳为中心,下至下界点、上至上界点的一个闭区间:我们就把这段时间作为可以匹配另一条流数据的“窗口”范围。所以对于另一条流(不妨叫B)中的数据元素b,如果它的时间戳落在了这个区间范围内,a和b就可以成功配对,进而进行计算输出结果。所以匹配的条件为:
a.timestamp + lowerBound <= b.timestamp <= a.timestamp + upperBound
这里需要注意,做间隔联结的两条流A和B,也必须基于相同的key;下界lowerBound应该小于等于上界upperBound,两者都可正可负;间隔联结目前只支持事件时间语义。
如下图所示,我们可以清楚地看到间隔联结的方式:
下方的流A去间隔联结上方的流B,所以基于A的每个数据元素,都可以开辟一个间隔区间。我们这里设置下界为-2毫秒,上界为1毫秒。于是对于时间戳为2的A中元素,它的可匹配区间就是[0, 3],流B中有时间戳为0、1的两个元素落在这个范围内,所以就可以得到匹配数据对(2, 0)和(2, 1)。同样地,A中时间戳为3的元素,可匹配区间为[1, 4],B中只有时间戳为1的一个数据可以匹配,于是得到匹配数据对(3, 1)。
所以我们可以看到,间隔联结同样是一种内连接(inner join)。与窗口联结不同的是,interval join做匹配的时间段是基于流中数据的,所以并不确定;而且流B中的数据可以不只在一个区间内被匹配。
间隔联结的调用
间隔联结在代码中,是基于KeyedStream的联结(join)操作。DataStream在keyBy得到KeyedStream之后,可以调用.intervalJoin()来合并两条流,传入的参数同样是一个KeyedStream,两者的key类型应该一致;得到的是一个IntervalJoin类型。后续的操作同样是完全固定的:先通过.between()方法指定间隔的上下界,再调用.process()方法,定义对匹配数据对的处理操作。调用.process()需要传入一个处理函数,这是处理函数家族的最后一员:“处理联结函数”ProcessJoinFunction。
通用调用形式
stream1
.keyBy(<KeySelector>)
.intervalJoin(stream2.keyBy(<KeySelector>))
.between(Time.milliseconds(-2), Time.milliseconds(1))
.process (new ProcessJoinFunction<Integer, Integer, String(){
@Override
public void processElement(Integer left, Integer right, Context ctx, Collector<String> out) {
out.collect(left + "," + right);
}
});
可以看到,抽象类ProcessJoinFunction就像是ProcessFunction和JoinFunction的结合,内部同样有一个抽象方法.processElement()。与其他处理函数不同的是,它多了一个参数,这自然是因为有来自两条流的数据。参数中left指的就是第一条流中的数据,right则是第二条流中与它匹配的数据。每当检测到一组匹配,就会调用这里的.processElement()方法,经处理转换之后输出结果。
间隔联结实例
案例需求:在电商网站中,某些用户行为往往会有短时间内的强关联。我们这里举一个例子,我们有两条流,一条是下订单的流,一条是支付的流。我们可以针对同一个订单,来做这样一个联结。也就是一个订单的和这个订单的支付数据进行一个联结查询。
代码实现:正常使用
package com.zxl.flink;
import com.zxl.bean.Orders;
import com.zxl.bean.Payments;
import com.zxl.datas.OrdersData;
import com.zxl.datas.PaymentData;
import org.apache.flink.api.common.eventtime.SerializableTimestampAssigner;
import org.apache.flink.api.common.eventtime.WatermarkStrategy;
import org.apache.flink.streaming.api.TimeCharacteristic;
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.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.functions.co.ProcessJoinFunction;
import org.apache.flink.streaming.api.windowing.time.Time;
import org.apache.flink.util.Collector;
import scala.Tuple9;
public class DemoTest {
public static void main(String[] args) throws Exception {
//创建Flink流处理执行环境
StreamExecutionEnvironment environment = StreamExecutionEnvironment.getExecutionEnvironment();
//设置并行度为1
environment.setParallelism(1);
// TODO: 2024/1/7 定义时间语义
environment.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);
//调用Flink自定义Source
// TODO: 2024/1/6 订单数据
DataStreamSource<Orders> ordersDataStreamSource = environment.addSource(new OrdersData());
// TODO: 2024/1/6 支付数据
DataStreamSource<Payments> paymentsDataStreamSource = environment.addSource(new PaymentData());
// TODO: 2024/1/7 配置订单数据水位线
SingleOutputStreamOperator<Orders> ordersWater = ordersDataStreamSource.assignTimestampsAndWatermarks(WatermarkStrategy
// TODO: 2024/1/7 指定watermark生成:升序的watermark,没有等待时间
.<Orders>forMonotonousTimestamps()
.withTimestampAssigner(new SerializableTimestampAssigner<Orders>() {
@Override
public long extractTimestamp(Orders orders, long l) {
return orders.getOrder_date();
}
})
);
// TODO: 2024/1/7 配置支付数据水位线
SingleOutputStreamOperator<Payments> paymentsWater = paymentsDataStreamSource.assignTimestampsAndWatermarks(WatermarkStrategy
.<Payments>forMonotonousTimestamps()
.withTimestampAssigner(new SerializableTimestampAssigner<Payments>() {
@Override
public long extractTimestamp(Payments payments, long l) {
return payments.getPayment_date();
}
})
);
// TODO interval join
//1. 分别做keyby,key其实就是关联条件
KeyedStream<Orders, Long> ordersKeyedStream = ordersWater.keyBy(orders -> orders.getOrder_id());
KeyedStream<Payments, Long> paymentsKeyedStream = paymentsWater.keyBy(payments -> payments.getOrder_id());
//2. 调用 interval join
SingleOutputStreamOperator<Tuple9> process = ordersKeyedStream.intervalJoin(paymentsKeyedStream)
.between(Time.seconds(-10), Time.seconds(10))
.process(new ProcessJoinFunction<Orders, Payments, Tuple9>() {
/**
* 两条流的数据匹配上,才会调用这个方法
* @param left ks1的数据
* @param right ks2的数据
* @param ctx 上下文
* @param out 采集器
* @throws Exception
*/
@Override
public void processElement(Orders orders, Payments payments, ProcessJoinFunction<Orders, Payments, Tuple9>.Context context, Collector<Tuple9> collector) throws Exception {
Tuple9<Long, Long, Long, Integer, Integer, Long, Integer, Long, String> tuple9 = new Tuple9<>(orders.getOrder_id(), orders.getUser_id(), orders.getOrder_date(), orders.getOrder_amount(), orders.getProduct_id(), orders.getOrder_num(), payments.getPayment_amount(), payments.getPayment_date(), payments.getPayment_type());
collector.collect(tuple9);
}
});
process.print();
environment.execute();
}
}
代码实现:处理迟到数据
package com.zxl.flink;
import com.zxl.bean.Orders;
import com.zxl.bean.Payments;
import com.zxl.datas.OrdersData;
import com.zxl.datas.PaymentData;
import org.apache.flink.api.common.eventtime.SerializableTimestampAssigner;
import org.apache.flink.api.common.eventtime.WatermarkStrategy;
import org.apache.flink.streaming.api.TimeCharacteristic;
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.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.functions.co.ProcessJoinFunction;
import org.apache.flink.streaming.api.windowing.time.Time;
import org.apache.flink.api.common.typeinfo.Types;
import org.apache.flink.util.Collector;
import org.apache.flink.util.OutputTag;
import scala.Tuple9;
import java.time.Duration;
public class DemoTest {
public static void main(String[] args) throws Exception {
//创建Flink流处理执行环境
StreamExecutionEnvironment environment = StreamExecutionEnvironment.getExecutionEnvironment();
//设置并行度为1
environment.setParallelism(1);
// TODO: 2024/1/7 定义时间语义
environment.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);
//调用Flink自定义Source
// TODO: 2024/1/6 订单数据
DataStreamSource<Orders> ordersDataStreamSource = environment.addSource(new OrdersData());
// TODO: 2024/1/6 支付数据
DataStreamSource<Payments> paymentsDataStreamSource = environment.addSource(new PaymentData());
// TODO: 2024/1/7 配置订单数据水位线
SingleOutputStreamOperator<Orders> ordersWater = ordersDataStreamSource.assignTimestampsAndWatermarks(WatermarkStrategy
// TODO: 2024/1/7 指定watermark生成:升序的watermark,没有等待时间
.<Orders>forBoundedOutOfOrderness(Duration.ofSeconds(3))
.withTimestampAssigner(new SerializableTimestampAssigner<Orders>() {
@Override
public long extractTimestamp(Orders orders, long l) {
return orders.getOrder_date() * 1000;
}
})
);
// TODO: 2024/1/7 配置支付数据水位线
SingleOutputStreamOperator<Payments> paymentsWater = paymentsDataStreamSource.assignTimestampsAndWatermarks(WatermarkStrategy
.<Payments>forBoundedOutOfOrderness(Duration.ofSeconds(3))
.withTimestampAssigner(new SerializableTimestampAssigner<Payments>() {
@Override
public long extractTimestamp(Payments payments, long l) {
return payments.getPayment_date() * 1000;
}
})
);
// TODO interval join
//配置测输出流标签
OutputTag<Orders> orders_tag = new OutputTag<Orders>("orders", Types.POJO(Orders.class));
OutputTag<Payments> payments_tag = new OutputTag<Payments>("payments", Types.POJO(Payments.class));
//1. 分别做keyby,key其实就是关联条件
KeyedStream<Orders, Long> ordersKeyedStream = ordersWater.keyBy(orders -> orders.getOrder_id());
KeyedStream<Payments, Long> paymentsKeyedStream = paymentsWater.keyBy(payments -> payments.getOrder_id());
//2. 调用 interval join
SingleOutputStreamOperator<Tuple9> process = ordersKeyedStream.intervalJoin(paymentsKeyedStream)
.between(Time.seconds(-2), Time.seconds(2))
.sideOutputLeftLateData(orders_tag)
.sideOutputRightLateData(payments_tag)
.process(new ProcessJoinFunction<Orders, Payments, Tuple9>() {
/**
* 两条流的数据匹配上,才会调用这个方法
* @param left ks1的数据
* @param right ks2的数据
* @param ctx 上下文
* @param out 采集器
* @throws Exception
*/
@Override
public void processElement(Orders orders, Payments payments, ProcessJoinFunction<Orders, Payments, Tuple9>.Context context, Collector<Tuple9> collector) throws Exception {
Tuple9<Long, Long, Long, Integer, Integer, Long, Integer, Long, String> tuple9 = new Tuple9<>(orders.getOrder_id(), orders.getUser_id(), orders.getOrder_date(), orders.getOrder_amount(), orders.getProduct_id(), orders.getOrder_num(), payments.getPayment_amount(), payments.getPayment_date(), payments.getPayment_type());
collector.collect(tuple9);
}
});
ordersKeyedStream.print();
paymentsKeyedStream.print();
process.print("主流");
process.getSideOutput(orders_tag).printToErr("orders迟到的数据");
process.getSideOutput(payments_tag).printToErr("payments迟到的数据");
environment.execute();
}
}