Apache Flink: Skewed data distribution on KeyedStream

问题

I have this Java code in Flink:

env.setParallelism(6);

//Read from Kafka topic with 12 partitions
DataStream<String> line = env.addSource(myConsumer);

//Filter half of the records 
DataStream<Tuple2<String, Integer>> line_Num_Odd = line_Num.filter(new FilterOdd());
DataStream<Tuple3<String, String, Integer>> line_Num_Odd_2 = line_Num_Odd.map(new OddAdder());

//Filter the other half
DataStream<Tuple2<String, Integer>> line_Num_Even = line_Num.filter(new FilterEven());
DataStream<Tuple3<String, String, Integer>> line_Num_Even_2 = line_Num_Even.map(new EvenAdder());

//Join all the data again
DataStream<Tuple3<String, String, Integer>> line_Num_U = line_Num_Odd_2.union(line_Num_Even_2);

//Window
DataStream<Tuple3<String, String, Integer>> windowedLine_Num_U_K = line_Num_U
                .keyBy(1)
                .window(TumblingProcessingTimeWindows.of(Time.seconds(10)))
                .reduce(new Reducer());

The problem is that the window should be able to process with parallelism = 2 as there are two diferent groups of data with keys "odd" and "even" in the second String in the Tuple3. Everything is running with parallelism 6 but not the window which is running with parallelism = 1 and I just need it to have parallelism = 2 because of my requirements.

The functions used in the code are the following:

public static class FilterOdd implements FilterFunction<Tuple2<String, Integer>> {

    public boolean filter(Tuple2<String, Integer> line) throws Exception {
        Boolean isOdd = (Long.valueOf(line.f0.split(" ")[0]) % 2) != 0;
        return isOdd;
    }
};


public static class FilterEven implements FilterFunction<Tuple2<String, Integer>> {

    public boolean filter(Tuple2<String, Integer> line) throws Exception {
        Boolean isEven = (Long.valueOf(line.f0.split(" ")[0]) % 2) == 0;
        return isEven;
    }
};

public static class OddAdder implements MapFunction<Tuple2<String, Integer>, Tuple3<String, String, Integer>> {

    public Tuple3<String, String, Integer> map(Tuple2<String, Integer> line) throws Exception {
        Tuple3<String, String, Integer> newLine = new Tuple3<String, String, Integer>(line.f0, "odd", line.f1);
        return newLine;
    }
};


public static class EvenAdder implements MapFunction<Tuple2<String, Integer>, Tuple3<String, String, Integer>> {

    public Tuple3<String, String, Integer> map(Tuple2<String, Integer> line) throws Exception {
        Tuple3<String, String, Integer> newLine = new Tuple3<String, String, Integer>(line.f0, "even", line.f1);
        return newLine;
    }
};

public static class Reducer implements ReduceFunction<Tuple3<String, String, Integer>> {

    public Tuple3<String, String, Integer> reduce(Tuple3<String, String, Integer> line1,
            Tuple3<String, String, Integer> line2) throws Exception {
        Long sum = Long.valueOf(line1.f0.split(" ")[0]) + Long.valueOf(line2.f0.split(" ")[0]);
        Long sumTS = Long.valueOf(line1.f0.split(" ")[1]) + Long.valueOf(line2.f0.split(" ")[1]);
        Tuple3<String, String, Integer> newLine = new Tuple3<String, String, Integer>(String.valueOf(sum) +
                " " + String.valueOf(sumTS), line1.f1, line1.f2 + line2.f2);
        return newLine;
    }
};

Thanks for your help!

SOLUTION: I have changed the content of the keys from "odd" and "even" to "odd0000" and "even1111" and it is working properly now.

回答1:

Keys are distributed to worker threads by hash partitioning. This means that the key values are hashed and the thread is determined by modulo #workers. With two keys and two threads there is a good chance that both keys are assigned to the same thread.

You can try to use different key values whose hash values distribute across both threads.

来源：https://stackoverflow.com/questions/44710398/apache-flink-skewed-data-distribution-on-keyedstream