Flink中的双流join和维表join

一、双流join

在数据库中的静态表上做OLAP分析时，两表join是非常常见的操作。同理，在流式处理作业中，有时也需要在两条流上做join以获得更丰富的信息。

1、Tumbling Window Join

代码示例：

DataStream<Integer> orangeStream = ...
DataStream<Integer> greenStream = ...

orangeStream.join(greenStream)
    .where(<KeySelector>)
    .equalTo(<KeySelector>)
    .window(TumblingEventTimeWindows.of(Time.milliseconds(2)))
    .apply(new JoinFunction<Integer, Integer, String> (){
        @Override
        public String join(Integer first, Integer second) {
            return first + "," + second;
        }
    });

2、Sliding Window Join

示例代码：

DataStream<Integer> orangeStream = ...
DataStream<Integer> greenStream = ...

orangeStream.join(greenStream)
    .where(<KeySelector>)
    .equalTo(<KeySelector>)
    .window(SlidingEventTimeWindows.of(Time.milliseconds(2) /* size */, Time.milliseconds(1) /* slide */))
    .apply(new JoinFunction<Integer, Integer, String> (){
        @Override
        public String join(Integer first, Integer second) {
            return first + "," + second;
        }
    });

3、Session Window Join

示例代码：

DataStream<Integer> orangeStream = ...
DataStream<Integer> greenStream = ...

orangeStream.join(greenStream)
    .where(<KeySelector>)
    .equalTo(<KeySelector>)
    .window(EventTimeSessionWindows.withGap(Time.milliseconds(1)))
    .apply(new JoinFunction<Integer, Integer, String> (){
        @Override
        public String join(Integer first, Integer second) {
            return first + "," + second;
        }
    });

以上3种都是“inner join”，只是窗口类型不一样。

4、Interval Join

右流相对左流偏移的时间区间进行关联，即：

right.timestamp ∈ [left.timestamp + lowerBound; left.timestamp + upperBound]

In the example above, we join two streams ‘orange’ and ‘green’ with a lower bound of -2 milliseconds and an upper bound of +1 millisecond. Be default, these boundaries are inclusive, but .lowerBoundExclusive() and .upperBoundExclusive can be applied to change the behaviour.

Using the more formal notation again this will translate to

orangeElem.ts + lowerBound <= greenElem.ts <= orangeElem.ts + upperBound

注意：目前 interval join 只支持 Event time

示例代码：

DataStream<Integer> orangeStream = ...
DataStream<Integer> greenStream = ...

orangeStream
    .keyBy(<KeySelector>)
    .intervalJoin(greenStream.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(first + "," + second);
        }
    });

5、coGroup

只有inner join肯定还不够，如何实现left/right outer join呢？答案就是利用coGroup()算子。它的调用方式类似于join()算子，也需要开窗，但是CoGroupFunction比JoinFunction更加灵活，可以按照用户指定的逻辑匹配左流和/或右流的数据并输出。

以下的例子就实现了点击流left join订单流的功能，是很朴素的nested loop join思想（二重循环）。

clickRecordStream
  .coGroup(orderRecordStream)
  .where(record -> record.getMerchandiseId())
  .equalTo(record -> record.getMerchandiseId())
  .window(TumblingProcessingTimeWindows.of(Time.seconds(10)))
  .apply(new CoGroupFunction<AnalyticsAccessLogRecord, OrderDoneLogRecord, Tuple2<String, Long>>() {
    @Override
    public void coGroup(Iterable<AnalyticsAccessLogRecord> accessRecords, Iterable<OrderDoneLogRecord> orderRecords, 
                        Collector<Tuple2<String, Long>> collector) throws Exception {
      for (AnalyticsAccessLogRecord accessRecord : accessRecords) {
        boolean isMatched = false;
        for (OrderDoneLogRecord orderRecord : orderRecords) {
          // 右流中有对应的记录
          collector.collect(new Tuple2<>(accessRecord.getMerchandiseName(), orderRecord.getPrice()));
          isMatched = true;
        }
        if (!isMatched) {
          // 右流中没有对应的记录
          collector.collect(new Tuple2<>(accessRecord.getMerchandiseName(), null));
        }
      }
    }
  })
  .print().setParallelism(1);

二、维表join

1、预加载维表

通过定义一个类实现RichMapFunction，在open()中读取维表数据加载到内存中，在probe流map()方法中与维表数据进行关联。

RichMapFunction中open方法里加载维表数据到内存的方式特点如下：

优点：实现简单
缺点：因为数据存于内存，所以只适合小数据量并且维表数据更新频率不高的情况下。虽然可以在open中定义一个定时器定时更新维表，但是还是存在维表更新不及时的情况。

class MapJoinDemo1 extends RichMapFunction<Tuple2<String, Integer>, Tuple3<String, Integer, String>> {
        //定义一个变量，用于保存维表数据在内存
        Map<Integer, String> dim;

        @Override
        public void open(Configuration parameters) throws Exception {
            //在open方法中读取维表数据，可以从数据中读取、文件中读取、接口中读取等等。
            dim = new HashMap<>();
            dim.put(1001, "beijing");
            dim.put(1002, "shanghai");
            dim.put(1003, "wuhan");
            dim.put(1004, "changsha");
        }

        @Override
        public Tuple3<String, Integer, String> map(Tuple2<String, Integer> value) throws Exception {
            //在map方法中进行主流和维表的关联
            String cityName = "";
            if (dim.containsKey(value.f1)) {
                cityName = dim.get(value.f1);
            }
            return new Tuple3<>(value.f0, value.f1, cityName);
        }
    }
}

2、热存储维表

这种方式是将维表数据存储在Redis、HBase、MySQL等外部存储中，实时流在关联维表数据的时候实时去外部存储中查询，这种方式特点如下：

优点：维度数据量不受内存限制，可以存储很大的数据量。
缺点：因为维表数据在外部存储中，读取速度受制于外部存储的读取速度；另外维表的同步也有延迟。

(1) 使用cache来减轻访问压力

可以使用缓存来存储一部分常访问的维表数据，以减少访问外部系统的次数，比如使用guava Cache。

class MapJoinDemo1 extends RichMapFunction<Tuple2<String, Integer>, Tuple3<String, Integer, String>> {
        LoadingCache<Integer, String> dim;

        @Override
        public void open(Configuration parameters) throws Exception {
            //使用google LoadingCache来进行缓存
            dim = CacheBuilder.newBuilder()
                    //最多缓存个数，超过了就根据最近最少使用算法来移除缓存
                    .maximumSize(1000)
                    //在更新后的指定时间后就回收
                    .expireAfterWrite(10, TimeUnit.MINUTES)
                    //指定移除通知
                    .removalListener(new RemovalListener<Integer, String>() {
                        @Override
                        public void onRemoval(RemovalNotification<Integer, String> removalNotification) {
                            System.out.println(removalNotification.getKey() + "被移除了，值为：" + removalNotification.getValue());
                        }
                    })
                    .build(
                            //指定加载缓存的逻辑
                            new CacheLoader<Integer, String>() {
                                @Override
                                public String load(Integer cityId) throws Exception {
                                    String cityName = readFromHbase(cityId);
                                    return cityName;
                                }
                            }
                    );
        }

        private String readFromHbase(Integer cityId) {
            //读取hbase，模拟从hbase读取数据
            Map<Integer, String> temp = new HashMap<>();
            temp.put(1001, "beijing");
            temp.put(1002, "shanghai");
            temp.put(1003, "wuhan");
            temp.put(1004, "changsha");
            String cityName = "";
            if (temp.containsKey(cityId)) {
                cityName = temp.get(cityId);
            }
            return cityName;
        }

        @Override
        public Tuple3<String, Integer, String> map(Tuple2<String, Integer> value) throws Exception {
            //在map方法中进行主流和维表的关联
            String cityName = "";
            if (dim.get(value.f1) != null) {
                cityName = dim.get(value.f1);
            }
            return new Tuple3<>(value.f0, value.f1, cityName);
        }
    }
}

(2) 使用异步IO来提高访问吞吐量

Flink与外部存储系统进行读写操作的时候可以使用同步方式，也就是发送一个请求后等待外部系统响应，然后再发送第二个读写请求，这样的方式吞吐量比较低，可以用提高并行度的方式来提高吞吐量，但是并行度多了也就导致了进程数量多了，占用了大量的资源。

Flink中可以使用异步IO来读写外部系统，这要求外部系统客户端支持异步IO，不过目前很多系统都支持异步IO客户端。但是如果使用异步就要涉及到三个问题：

超时：如果查询超时那么就认为是读写失败，需要按失败处理；
并发数量：如果并发数量太多，就要触发Flink的反压机制来抑制上游的写入。
返回顺序错乱：顺序错乱了要根据实际情况来处理，Flink支持两种方式：允许乱序、保证顺序。

public class JoinDemo3 {
    public static void main(String[] args) throws Exception {

        StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        DataStream<Tuple2<String, Integer>> textStream = env.socketTextStream("localhost", 9000, "\n")
                .map(p -> {
                    //输入格式为：user,1000,分别是用户名称和城市编号
                    String[] list = p.split(",");
                    return new Tuple2<String, Integer>(list[0], Integer.valueOf(list[1]));
                })
                .returns(new TypeHint<Tuple2<String, Integer>>() {
                });


        DataStream<Tuple3<String,Integer, String>> orderedResult = AsyncDataStream
                //保证顺序：异步返回的结果保证顺序，超时时间1秒，最大容量2，超出容量触发反压
                .orderedWait(textStream, new JoinDemo3AyncFunction(), 1000L, TimeUnit.MILLISECONDS, 2)
                .setParallelism(1);

        DataStream<Tuple3<String,Integer, String>> unorderedResult = AsyncDataStream
                //允许乱序：异步返回的结果允许乱序，超时时间1秒，最大容量2，超出容量触发反压
                .unorderedWait(textStream, new JoinDemo3AyncFunction(), 1000L, TimeUnit.MILLISECONDS, 2)
                .setParallelism(1);

        orderedResult.print();
        unorderedResult.print();
        env.execute("joinDemo");
    }

    //定义个类，继承RichAsyncFunction，实现异步查询存储在mysql里的维表
    //输入用户名、城市ID，返回 Tuple3<用户名、城市ID，城市名称>
    static class JoinDemo3AyncFunction extends RichAsyncFunction<Tuple2<String, Integer>, Tuple3<String, Integer, String>> {
        // 链接
        private static String jdbcUrl = "jdbc:mysql://192.168.145.1:3306?useSSL=false";
        private static String username = "root";
        private static String password = "123";
        private static String driverName = "com.mysql.jdbc.Driver";
        java.sql.Connection conn;
        PreparedStatement ps;

        @Override
        public void open(Configuration parameters) throws Exception {
            super.open(parameters);

            Class.forName(driverName);
            conn = DriverManager.getConnection(jdbcUrl, username, password);
            ps = conn.prepareStatement("select city_name from tmp.city_info where id = ?");
        }

        @Override
        public void close() throws Exception {
            super.close();
            conn.close();
        }

        //异步查询方法
        @Override
        public void asyncInvoke(Tuple2<String, Integer> input, ResultFuture<Tuple3<String,Integer, String>> resultFuture) throws Exception {
            // 使用 city id 查询
            ps.setInt(1, input.f1);
            ResultSet rs = ps.executeQuery();
            String cityName = null;
            if (rs.next()) {
                cityName = rs.getString(1);
            }
            List list = new ArrayList<Tuple2<Integer, String>>();
            list.add(new Tuple3<>(input.f0,input.f1, cityName));
            resultFuture.complete(list);
        }

        //超时处理
        @Override
        public void timeout(Tuple2<String, Integer> input, ResultFuture<Tuple3<String,Integer, String>> resultFuture) throws Exception {
            List list = new ArrayList<Tuple2<Integer, String>>();
            list.add(new Tuple3<>(input.f0,input.f1, ""));
            resultFuture.complete(list);
        }
    }
}

3、广播维表

利用Flink的Broadcast State将维度数据流广播到下游做join操作。特点如下：

优点：维度数据变更后可以即时更新到结果中。
缺点：数据保存在内存中，支持的维度数据量比较小。

        //定义城市流
        DataStream<Tuple2<Integer, String>> cityStream = env.socketTextStream("localhost", 9001, "\n")
                .map(p -> {
                    //输入格式为：城市ID,城市名称
                    String[] list = p.split(",");
                    return new Tuple2<Integer, String>(Integer.valueOf(list[0]), list[1]);
                })
                .returns(new TypeHint<Tuple2<Integer, String>>() {
                });

        //将城市流定义为广播流
        final MapStateDescriptor<Integer, String> broadcastDesc = new MapStateDescriptor("broad1", Integer.class, String.class);
        BroadcastStream<Tuple2<Integer, String>> broadcastStream = cityStream.broadcast(broadcastDesc);

        DataStream result = textStream.connect(broadcastStream)
                .process(new BroadcastProcessFunction<Tuple2<String, Integer>, Tuple2<Integer, String>, Tuple3<String, Integer, String>>() {
                    //处理非广播流，关联维度
                    @Override
                    public void processElement(Tuple2<String, Integer> value, ReadOnlyContext ctx, Collector<Tuple3<String, Integer, String>> out) throws Exception {
                        ReadOnlyBroadcastState<Integer, String> state = ctx.getBroadcastState(broadcastDesc);
                        String cityName = "";
                        if (state.contains(value.f1)) {
                            cityName = state.get(value.f1);
                        }
                        out.collect(new Tuple3<>(value.f0, value.f1, cityName));
                    }

                    @Override
                    public void processBroadcastElement(Tuple2<Integer, String> value, Context ctx, Collector<Tuple3<String, Integer, String>> out) throws Exception {
                        System.out.println("收到广播数据：" + value);
                        ctx.getBroadcastState(broadcastDesc).put(value.f0, value.f1);
                    }
                });

4、 Temporal table function join

Temporal table是持续变化表上某一时刻的视图，Temporal table function是一个表函数，传递一个时间参数，返回Temporal table这一指定时刻的视图。

可以将维度数据流映射为Temporal table，主流与这个Temporal table进行关联，可以关联到某一个版本（历史上某一个时刻）的维度数据。

Temporal table function join的特点如下：

优点：维度数据量可以很大，维度数据更新及时，不依赖外部存储，可以关联不同版本的维度数据。
缺点：只支持在Flink SQL API中使用。

public class JoinDemo5 {
    public static void main(String[] args) throws Exception {
        StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        EnvironmentSettings bsSettings = EnvironmentSettings.newInstance().useBlinkPlanner().inStreamingMode().build();
        StreamTableEnvironment tableEnv = StreamTableEnvironment.create(env, bsSettings);

        //定义主流
        DataStream<Tuple2<String, Integer>> textStream = env.socketTextStream("localhost", 9000, "\n")
                .map(p -> {
                    //输入格式为：user,1000,分别是用户名称和城市编号
                    String[] list = p.split(",");
                    return new Tuple2<String, Integer>(list[0], Integer.valueOf(list[1]));
                })
                .returns(new TypeHint<Tuple2<String, Integer>>() {
                });

        //定义城市流
        DataStream<Tuple2<Integer, String>> cityStream = env.socketTextStream("localhost", 9001, "\n")
                .map(p -> {
                    //输入格式为：城市ID,城市名称
                    String[] list = p.split(",");
                    return new Tuple2<Integer, String>(Integer.valueOf(list[0]), list[1]);
                })
                .returns(new TypeHint<Tuple2<Integer, String>>() {
                });

        //转变为Table
        Table userTable = tableEnv.fromDataStream(textStream, "user_name,city_id,ps.proctime");
        Table cityTable = tableEnv.fromDataStream(cityStream, "city_id,city_name,ps.proctime");

        //定义一个TemporalTableFunction
        TemporalTableFunction dimCity = cityTable.createTemporalTableFunction("ps", "city_id");
        //注册表函数
        tableEnv.registerFunction("dimCity", dimCity);

        //关联查询
        Table result = tableEnv
                .sqlQuery("select u.user_name,u.city_id,d.city_name from " + userTable + " as u " +
                        ", Lateral table (dimCity(u.ps)) d " +
                        "where u.city_id=d.city_id");
        
        //打印输出
        DataStream resultDs = tableEnv.toAppendStream(result, Row.class);
        resultDs.print();
        env.execute("joinDemo");
    }
}

5、四种维表关联方式比较

参考：

https://www.jianshu.com/p/45ec888332df

https://ci.apache.org/projects/flink/flink-docs-release-1.12/dev/stream/operators/joining.html

https://blog.csdn.net/chybin500/article/details/106482620

posted @ 2021-01-07 11:46 阿凡卢阅读(4744) 评论(0) 编辑收藏举报

刷新页面返回顶部

Loading

阿凡卢

If you never try, you will never know.

Flink中的双流join和维表join

一、双流join

1、Tumbling Window Join

2、Sliding Window Join

3、Session Window Join

4、Interval Join

5、coGroup

二、维表join

1、预加载维表

2、热存储维表

(1) 使用cache来减轻访问压力

(2) 使用异步IO来提高访问吞吐量

3、广播维表

4、 Temporal table function join

5、四种维表关联方式比较

公告

Loading

阿凡卢

If you never try, you will never know.

Flink中的双流join和维表join

一、双流join

1、Tumbling Window Join

2、Sliding Window Join

3、Session Window Join

4、Interval Join

5、coGroup

二、维表join

1、 预加载维表

2、 热存储维表

(1) 使用cache来减轻访问压力

(2) 使用异步IO来提高访问吞吐量

3、 广播维表

4、 Temporal table function join

5、四种维表关联方式比较

公告

1、预加载维表

2、热存储维表

3、广播维表