上一篇 Apache Flink 学习笔记（一） 简单示范了批处理的使用，本篇展示流式处理的使用方法。
流处理也叫无界处理，因为数据是源源不断的被加载进来的，流处理需要用到DataStream类。本篇demo 将结合kafka（公司有现成的消息生产者）来演示。
kafka 消息体如下（json）：
{
"appId":"xxxx",
"module":"xxxx"
//其余省略
}
现在我想每10s统计一次，按照appid分组计数（需求简单一点），Event Time为ProcessingTime，Windows为滚动窗口。
import com.alibaba.fastjson.JSON;
import com.alibaba.fastjson.JSONObject;
import org.apache.flink.api.common.functions.AggregateFunction;
import org.apache.flink.api.common.functions.MapFunction;
import org.apache.flink.api.java.tuple.Tuple2;
import org.apache.flink.streaming.api.TimeCharacteristic;
import org.apache.flink.streaming.api.datastream.DataStream;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.functions.sink.SinkFunction;
import org.apache.flink.streaming.api.windowing.time.Time;
import org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumer09;
import java.util.Date;
import java.util.HashMap;
import java.util.Map;
public class Demo3 {
public static void main(String[] args) {
//生成流式执行环境对象 StreamExecutionEnvironment
final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
env.getConfig().enableSysoutLogging();//开启Sysout打日志
env.setStreamTimeCharacteristic(TimeCharacteristic.ProcessingTime); //设置窗口的时间单位为process time
env.setParallelism(2);//全局并发数
//配置kafka bootstrap.servers
Properties properties = new Properties();
properties.setProperty("bootstrap.servers", "kafka bootstrap.servers");
//配置消息主题和应用名（自定义工具类FlinkKafkaManager，源码在后面）
FlinkKafkaManager manager = new FlinkKafkaManager("kafka.topic", "app.name", properties);
//用JsonObject 反序列化接收kafka
FlinkKafkaConsumer09<JSONObject> consumer = manager.build(JSONObject.class);
//从最新的消息开始接收
consumer.setStartFromLatest();
//获得DataStream
DataStream<JSONObject> messageStream = env.addSource(consumer);
//转化为pojo
DataStream<Bean3> bean3DataStream = messageStream.map(new FlatMap());
bean3DataStream
.keyBy(Bean3::getAppId) //也可以用“appId”替换
.timeWindow(Time.seconds(10))//等价于下面这一行，因为上面设置了TimeCharacteristic.ProcessingTime
// .window(TumblingProcessingTimeWindows.of(Time.seconds(10)))//基于process time的窗口
.aggregate(new Agg()) //聚合函数，这里也可以参照demo2用reduce函数
.addSink(new Sink()); //输出函数
try {
env.execute("app.name");//流式处理需要调用触发
} catch (Exception e) {
e.printStackTrace();
}
}
public static class FlatMap implements MapFunction<JSONObject, Bean3> {
@Override
public Bean3 map(JSONObject jsonObject) throws Exception {
return new Bean3(jsonObject.getString("appId"), jsonObject.getString("module"));
}
}
public static class Agg implements AggregateFunction<Bean3, Tuple2<Bean3, Long>, Tuple2<Bean3, Long>> {
@Override
public Tuple2<Bean3, Long> createAccumulator() {
return new Tuple2<Bean3, Long>();
}
@Override
public Tuple2<Bean3, Long> add(Bean3 bean3, Tuple2<Bean3, Long> bean3LongTuple2) {
Bean3 bean = bean3LongTuple2.f0;
Long count = bean3LongTuple2.f1;
if (bean == null) {
bean = bean3;
}
if (count == null) {
count = 1L;
} else {
count++;
}
return new Tuple2<>(bean, count);
}
@Override
public Tuple2<Bean3, Long> getResult(Tuple2<Bean3, Long> bean3LongTuple2) {
return bean3LongTuple2;
}
@Override
public Tuple2<Bean3, Long> merge(Tuple2<Bean3, Long> bean3LongTuple2, Tuple2<Bean3, Long> acc1) {
Bean3 bean = bean3LongTuple2.f0;
Long count = bean3LongTuple2.f1;
Long acc = acc1.f1;
return new Tuple2<>(bean, count + acc);
}
}
public static class Sink implements SinkFunction<Tuple2<Bean3, Long>> {
@Override
public void invoke(Tuple2<Bean3, Long> value, Context context) throws Exception {
System.out.println(value.f0.toString() + "," + value.f1);
}
}
public static class Bean3 {
public String appId;
public String module;
public Bean3() {
}
public Bean3(String appId, String module) {
this.appId = appId;
this.module = module;
}
public String getAppId() {
return appId;
}
public void setAppId(String appId) {
this.appId = appId;
}
public String getModule() {
return module;
}
public void setModule(String module) {
this.module = module;
}
@Override
public String toString() {
return "Bean3{" +
"appId='" + appId + '\'' +
", module='" + module + '\'' +
'}';
}
}
}
与上一篇批处理的demo相比，流处理显得复杂了许多。实际上二者有很多想通的地方，比如批处理中的groupBy和流处理的keyBy，都是按照指定维度分组的。
而流处理中会引入窗口的概念，正如前面所说，流式数据是无界数据，Flink 借助窗口将无界数据转化成一个个“批处理”再做计算。窗口分为滚动窗口，滑动窗口，会话窗口等等，具体可参见官网介绍。而每个窗口的时间划分则是由event time 决定的，本例采用的是ProcessingTime即处理时间。
下面我将demo3改造，使其变成使用EventTime，也就是说窗口的时间由数据源的时间戳（事件发生）决定。
改动1
//为pojo Bean3 添加时间戳字段
public static class Bean3 {
public Long timestamp;//add event time
public String appId;
public String module;
public Bean3() {
}
public Bean3(Long timestamp, String appId, String module) {
this.timestamp = timestamp;
this.appId = appId;
this.module = module;
}
public long getTimestamp() {
return timestamp;
}
public void setTimestamp(Long timestamp) {
this.timestamp = timestamp;
}
//省略其他
}
改动2
//设置窗口的时间单位为event time
env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);
改动3
//新增
//指定数据源的时间戳，Time.seconds(int)是指允许多长时间消息延迟
DataStream<Bean3> bean3DataStreamWithAssignTime =
bean3DataStream.assignTimestampsAndWatermarks(new BoundedOutOfOrdernessTimestampExtractor<Bean3>(Time.seconds(0)) {
@Override
public long extractTimestamp(Bean3 element) {
return element.getTimestamp();
}
});
改动4
bean3DataStreamWithAssignTime
.keyBy(Bean3::getAppId)
.window(TumblingEventTimeWindows.of(Time.seconds(10)))//基于event time的窗口
.allowedLateness(Time.seconds(5)) //允许数据延迟多长时间,谨慎使用,迟到的数据会导致出现重复
//后面省略
FlinkKafkaManager 源码
package flink.test.manager;
import org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumer09;
import java.util.Properties;
public class FlinkKafkaManager<T> {
private String topic;
private String groupId;
private Properties properties;
public FlinkKafkaManager(String topic, String groupId, Properties properties) {
this.topic = topic;
this.groupId = groupId;
this.properties = properties;
this.properties.setProperty("group.id", this.groupId);
//为使用默认kafka的用户配置基础配置
this.setDefaultKafkaProperties();
}
private void setDefaultKafkaProperties() {
//启用auto commit offset, 每5s commit一次
this.properties.setProperty("enable.auto.commit", "true");
this.properties.setProperty("auto.commit.interval.ms", "5000");
}
public FlinkKafkaConsumer09<T> build(Class<T> clazz) {
if (checkProperties()) {
return new FlinkKafkaConsumer09<T>(topic, new ConsumerDeserializationSchema(clazz), properties);
} else {
return null;
}
}
private boolean checkProperties() {
boolean isValued = true;
if (!properties.containsKey("bootstrap.servers")) {
isValued = false;
} else {
String brokers = properties.getProperty("bootstrap.servers");
if (brokers == null || brokers.isEmpty()) {
isValued = false;
}
}
if (this.topic == null || this.topic.isEmpty()) {
isValued = false;
}
if (!properties.containsKey("group.id")) {
isValued = false;
} else {
String groupId = properties.getProperty("group.id");
if (groupId == null || groupId.isEmpty()) {
isValued = false;
}
}
return isValued;
}
}
ConsumerDeserializationSchema 源码
package flink.test.manager;
import com.alibaba.fastjson.JSON;
import org.apache.flink.api.common.serialization.DeserializationSchema;
import org.apache.flink.api.common.typeinfo.TypeInformation;
import org.apache.flink.api.java.typeutils.TypeExtractor;
import java.io.IOException;
public class ConsumerDeserializationSchema<T> implements DeserializationSchema<T> {
private Class<T> clazz;
public ConsumerDeserializationSchema(Class<T> clazz) {
this.clazz = clazz;
}
@Override
public T deserialize(byte[] bytes) throws IOException {
//确保 new String(bytes) 是json 格式，如果不是，请自行解析
return JSON.parseObject(new String(bytes), clazz);
}
@Override
public boolean isEndOfStream(T t) {
return false;
}
@Override
public TypeInformation<T> getProducedType() {
return TypeExtractor.getForClass(clazz);
}
}
网友评论