重新认识mapreduce

写这篇文章，是因为最近遇到了mapreduce的二次排序问题。以前的理解不完全正确。首先看一下mapreduce的过程

相信这张图熟悉MR的人都应该见过，再来一张图

wordcount也不细说了，hadoop里面的hello,world

之前我的理解是map过来的<k,v>会形成(k,<v1,v2,v3...>)的格式，并且按照这种思路写出来不少的mapreduce程序，而且没有错。

后来自定义Writable对象，封装一组值作为key，也没有什么问题，而且一直认为key只要在compareTo中重写 了方法就万事大吉，而且compareTo返回0的会作为相同的key。误区就在这里，之前一直认为key相同的value会合并到一个"list"中-。这句话就有错，key是key，value是value，根本不会将key对应的value合并在一起，真实情况是默认将key相同（compareTo返回0的）的合并成了一组，在组相同的里面去foreach里面的value，如果是自定义key的话你可以将key打印一下，或发现key并不相同。

上代码：

public class Entry implements WritableComparable<Entry> {
    private String yearMonth;
    private int count;

    public Entry() {
    }

    @Override
    public int compareTo(Entry entry) {
        int result = this.yearMonth.compareTo(entry.getYearMonth());
        if (result == 0) {
            result = Integer.compare(count, entry.getCount());
        }
        return result;
    }

    @Override
    public void write(DataOutput dataOutput) throws IOException {
        dataOutput.writeUTF(yearMonth);
        dataOutput.writeInt(count);
    }

    @Override
    public void readFields(DataInput dataInput) throws IOException {
        this.yearMonth = dataInput.readUTF();
        this.count = dataInput.readInt();
    }

    public String getYearMonth() {
        return yearMonth;
    }

    public void setYearMonth(String yearMonth) {
        this.yearMonth = yearMonth;
    }

    public int getCount() {
        return count;
    }

    public void setCount(int count) {
        this.count = count;
    }

    @Override
    public String toString() {
        return yearMonth;
    }
}

自定义分区 EntryPartitioner.java

public class EntryPartitioner extends Partitioner<Entry, Text> {

    @Override
    public int getPartition(Entry entry, Text paramVALUE, int numberPartitions) {
        return Math.abs((entry.getYearMonth().hashCode() % numberPartitions));
    }
}

自定义分组

public class EntryGroupingComparator extends WritableComparator {
    public EntryGroupingComparator() {
        super(Entry.class, true);
    }

    @Override
    public int compare(WritableComparable a, WritableComparable b) {
        Entry a1 = (Entry) a;
        Entry b1 = (Entry) b;
        return a1.getYearMonth().compareTo(b1.getYearMonth());
    }
}

mapper类

public class SecondarySortMapper extends
        Mapper<LongWritable, Text, Entry, Text> {

    private Entry entry = new Entry();
    private Text value = new Text();

    @Override
    protected void map(LongWritable key, Text lines, Context context)
            throws IOException, InterruptedException {
        String line = lines.toString();
        String[] tokens = line.split(",");
        String yearMonth = tokens[0] + "-" + tokens[1];
        int count = Integer.parseInt(tokens[2]);

        entry.setYearMonth(yearMonth);
        entry.setCount(count);
        value.set(tokens[2]);
        context.write(entry, value);

    }
}

reducer类

public class SecondarySortReducer extends Reducer<Entry, Text, Entry, Text> {
    @Override
    protected void reduce(Entry key, Iterable<Text> values, Context context)
            throws IOException, InterruptedException {
        System.out.println("-----------------华丽的分割线-----------------");
        StringBuilder builder = new StringBuilder();
        for (Text value : values) {
            System.out.println(key+"==>"+value);
            builder.append(value.toString());
            builder.append(",");
        }
        context.write(key, new Text(builder.toString()));
    }
}

reducer中打印出来的跟原来想的不一样，一组的值除了自定义分组的属性相同外，其他的属性有不同的。看来以前是自己理解不够深入啊，特此写出，以示警戒