Spark SQL 用户自定义函数UDF、用户自定义聚合函数UDAF 教程（Java踩坑教学版）

在Spark中，也支持Hive中的自定义函数。自定义函数大致可以分为三种：

• UDF(User-Defined-Function)，即最基本的自定义函数，类似to_char,to_date等
• UDAF（User- Defined Aggregation Funcation），用户自定义聚合函数，类似在group by之后使用的sum,avg等
• UDTF(User-Defined Table-Generating Functions),用户自定义生成函数，有点像stream里面的flatMap

本篇就手把手教你如何编写UDF和UDAF

先来个简单的UDF

场景：
我们有这样一个文本文件：

1^^d
2^b^d
3^c^d
4^^d

在读取数据的时候，第二列的数据如果为空，需要显示'null'，不为空就直接输出它的值。定义完成后，就可以直接在SparkSQL中使用了。

代码为：

package test;

import org.apache.spark.SparkConf;
import org.apache.spark.api.java.JavaRDD;
import org.apache.spark.api.java.JavaSparkContext;
import org.apache.spark.sql.DataFrame;
import org.apache.spark.sql.Row;
import org.apache.spark.sql.RowFactory;
import org.apache.spark.sql.SQLContext;
import org.apache.spark.sql.types.DataTypes;
import org.apache.spark.sql.types.StructField;
import org.apache.spark.sql.types.StructType;

import java.util.ArrayList;
import java.util.List;

/**
 * Created by xinghailong on 2017/2/23.
 */
public class test3 {
    public static void main(String[] args) {
        //创建spark的运行环境
        SparkConf sparkConf = new SparkConf();
        sparkConf.setMaster("local[2]");
        sparkConf.setAppName("test-udf");
        JavaSparkContext sc = new JavaSparkContext(sparkConf);
        SQLContext sqlContext = new SQLContext(sc);
        //注册自定义方法
        sqlContext.udf().register("isNull", (String field,String defaultValue)->field==null?defaultValue:field, DataTypes.StringType);
        //读取文件
        JavaRDD<String> lines = sc.textFile( "C:\\test-udf.txt" );
        JavaRDD<Row> rows = lines.map(line-> RowFactory.create(line.split("\\^")));

        List<StructField> structFields = new ArrayList<StructField>();
        structFields.add(DataTypes.createStructField( "a", DataTypes.StringType, true ));
        structFields.add(DataTypes.createStructField( "b", DataTypes.StringType, true ));
        structFields.add(DataTypes.createStructField( "c", DataTypes.StringType, true ));
        StructType structType = DataTypes.createStructType( structFields );

        DataFrame test = sqlContext.createDataFrame( rows, structType);
        test.registerTempTable("test");
        
        sqlContext.sql("SELECT con_join(c,b) FROM test GROUP BY a").show();
        sc.stop();
    }
}

输出内容为：

+---+----+---+
|  a| _c1|  c|
+---+----+---+
|  1|null|  d|
|  2|   b|  d|
|  3|   c|  d|
|  4|null|  d|
+---+----+---+

其中比较关键的就是这句:

sqlContext.udf().register("isNull", (String field,String defaultValue)->field==null?defaultValue:field, DataTypes.StringType);

这里我直接用的java8的语法写的，如果是java8之前的版本，需要使用Function2创建匿名函数。

再来个自定义的UDAF—求平均数

先来个最简单的UDAF，求平均数。类似这种的操作有很多，比如最大值，最小值，累加，拼接等等，都可以采用相同的思路来做。

首先是需要定义UDAF函数

package test;

import org.apache.spark.sql.Row;
import org.apache.spark.sql.expressions.MutableAggregationBuffer;
import org.apache.spark.sql.expressions.UserDefinedAggregateFunction;
import org.apache.spark.sql.types.DataType;
import org.apache.spark.sql.types.DataTypes;
import org.apache.spark.sql.types.StructField;
import org.apache.spark.sql.types.StructType;

import java.util.ArrayList;
import java.util.List;

/**
 * Created by xinghailong on 2017/2/23.
 */
public class MyAvg extends UserDefinedAggregateFunction {

    @Override
    public StructType inputSchema() {
        List<StructField> structFields = new ArrayList<>();
        structFields.add(DataTypes.createStructField( "field1", DataTypes.StringType, true ));
        return DataTypes.createStructType( structFields );
    }

    @Override
    public StructType bufferSchema() {
        List<StructField> structFields = new ArrayList<>();
        structFields.add(DataTypes.createStructField( "field1", DataTypes.IntegerType, true ));
        structFields.add(DataTypes.createStructField( "field2", DataTypes.IntegerType, true ));
        return DataTypes.createStructType( structFields );
    }

    @Override
    public DataType dataType() {
        return DataTypes.IntegerType;
    }

    @Override
    public boolean deterministic() {
        return false;
    }

    @Override
    public void initialize(MutableAggregationBuffer buffer) {
        buffer.update(0,0);
        buffer.update(1,0);
    }

    @Override
    public void update(MutableAggregationBuffer buffer, Row input) {
        buffer.update(0,buffer.getInt(0)+1);
        buffer.update(1,buffer.getInt(1)+Integer.valueOf(input.getString(0)));
    }

    @Override
    public void merge(MutableAggregationBuffer buffer1, Row buffer2) {
        buffer1.update(0,buffer1.getInt(0)+buffer2.getInt(0));
        buffer1.update(1,buffer1.getInt(1)+buffer2.getInt(1));
    }

    @Override
    public Object evaluate(Row buffer) {
        return buffer.getInt(1)/buffer.getInt(0);
    }
}

使用的时候，需要先注册，然后在spark sql里面就可以直接使用了:

package test;

import com.tgou.standford.misdw.udf.MyAvg;
import org.apache.spark.SparkConf;
import org.apache.spark.api.java.JavaRDD;
import org.apache.spark.api.java.JavaSparkContext;
import org.apache.spark.sql.DataFrame;
import org.apache.spark.sql.Row;
import org.apache.spark.sql.RowFactory;
import org.apache.spark.sql.SQLContext;
import org.apache.spark.sql.types.DataTypes;
import org.apache.spark.sql.types.StructField;
import org.apache.spark.sql.types.StructType;

import java.util.ArrayList;
import java.util.List;

/**
 * Created by xinghailong on 2017/2/23.
 */
public class test4 {
    public static void main(String[] args) {
        SparkConf sparkConf = new SparkConf();
        sparkConf.setMaster("local[2]");
        sparkConf.setAppName("test");
        JavaSparkContext sc = new JavaSparkContext(sparkConf);
        SQLContext sqlContext = new SQLContext(sc);

        sqlContext.udf().register("my_avg",new MyAvg());

        JavaRDD<String> lines = sc.textFile( "C:\\test4.txt" );
        JavaRDD<Row> rows = lines.map(line-> RowFactory.create(line.split("\\^")));

        List<StructField> structFields = new ArrayList<StructField>();
        structFields.add(DataTypes.createStructField( "a", DataTypes.StringType, true ));
        structFields.add(DataTypes.createStructField( "b", DataTypes.StringType, true ));
        StructType structType = DataTypes.createStructType( structFields );

        DataFrame test = sqlContext.createDataFrame( rows, structType);
        test.registerTempTable("test");

        sqlContext.sql("SELECT my_avg(b) FROM test GROUP BY a").show();

        sc.stop();
    }
}

计算的文本内容为：

a^3
a^6
b^2
b^4
b^6

再来个无所不能的UDAF

真正的业务场景里面，总会有千奇百怪的需求，比如：

• 想要按照某个字段分组，取其中的一个最大值
• 想要按照某个字段分组，对分组内容的数据按照特定字段统计累加
• 想要按照某个字段分组，针对特定的条件，拼接字符串

再比如一个场景，需要按照某个字段分组，然后分组内的数据，又需要按照某一列进行去重，最后再计算值

• 1 按照某个字段分组
• 2 分组校验条件
• 3 然后处理字段

如果不用UDAF，你要是写spark可能需要这样做:

rdd.groupBy(r->r.xxx)
    .map(t2->{
        HashSet<String> set = new HashSet<>();
        for(Object p : t2._2){
            if(p.getBs() > 0 ){
                map.put(xx,yyy)
            }
        }
        return StringUtils.join(set.toArray(),",");
    });

上面是一段伪码，不保证正常运行哈。

这样写，其实也能应付需求了，但是代码显得略有点丑陋。还是不如SparkSQL看的清晰明了...

所以我们再尝试用SparkSql中的UDAF来一版！

首先需要创建UDAF类

import org.apache.commons.lang.StringUtils;
import org.apache.spark.sql.Row;
import org.apache.spark.sql.expressions.MutableAggregationBuffer;
import org.apache.spark.sql.expressions.UserDefinedAggregateFunction;
import org.apache.spark.sql.types.*;

import java.util.*;

/**
 *
 * Created by xinghailong on 2017/2/23.
 */
public class ConditionJoinUDAF extends UserDefinedAggregateFunction {
    @Override
    public StructType inputSchema() {
        List<StructField> structFields = new ArrayList<>();
        structFields.add(DataTypes.createStructField( "field1", DataTypes.IntegerType, true ));
        structFields.add(DataTypes.createStructField( "field2", DataTypes.StringType, true ));
        return DataTypes.createStructType( structFields );
    }

    @Override
    public StructType bufferSchema() {
        List<StructField> structFields = new ArrayList<>();
        structFields.add(DataTypes.createStructField( "field", DataTypes.StringType, true ));
        return DataTypes.createStructType( structFields );
    }

    @Override
    public DataType dataType() {
        return DataTypes.StringType;
    }

    @Override
    public boolean deterministic() {//是否强制每次执行的结果相同
        return false;
    }

    @Override
    public void initialize(MutableAggregationBuffer buffer) {//初始化
        buffer.update(0,"");
    }

    @Override
    public void update(MutableAggregationBuffer buffer, Row input) {//相同的executor间的数据合并
        Integer bs = input.getInt(0);
        String field = buffer.getString(0);
        String in = input.getString(1);
        if(bs > 0 && !"".equals(in) && !field.contains(in)){
            field += ","+in;
        }
        buffer.update(0,field);
    }

    @Override
    public void merge(MutableAggregationBuffer buffer1, Row buffer2) {//不同excutor间的数据合并
        String field1 = buffer1.getString(0);
        String field2 = buffer2.getString(0);
        if(!"".equals(field2)){
            field1 += ","+field2;
        }
        buffer1.update(0,field1);
    }

    @Override
    public Object evaluate(Row buffer) {//根据Buffer计算结果
        return StringUtils.join(Arrays.stream(buffer.getString(0).split(",")).filter(line->!line.equals("")).toArray(),",");
    }
}

拿一个例子坐下实验：

a^1111^2
a^1111^2
a^1111^2
a^1111^2
a^1111^2
a^2222^0
a^3333^1
b^4444^0
b^5555^3
c^6666^0

按照第一列进行分组，不同的第三列值，进行拼接。

package test;

import test.ConditionJoinUDAF;
import org.apache.spark.SparkConf;
import org.apache.spark.api.java.JavaRDD;
import org.apache.spark.api.java.JavaSparkContext;
import org.apache.spark.sql.DataFrame;
import org.apache.spark.sql.Row;
import org.apache.spark.sql.RowFactory;
import org.apache.spark.sql.SQLContext;
import org.apache.spark.sql.types.DataTypes;
import org.apache.spark.sql.types.StructField;
import org.apache.spark.sql.types.StructType;

import java.util.ArrayList;
import java.util.List;

/**
 * Created by xinghailong on 2017/2/23.
 */
public class test2 {
    public static void main(String[] args) {
        SparkConf sparkConf = new SparkConf();
        sparkConf.setMaster("local[2]");
        sparkConf.setAppName("test");
        JavaSparkContext sc = new JavaSparkContext(sparkConf);
        SQLContext sqlContext = new SQLContext(sc);

        sqlContext.udf().register("con_join",new ConditionJoinUDAF());

        JavaRDD<String> lines = sc.textFile( "C:\\test2.txt" );
        JavaRDD<Row> rows = lines.map(line-> RowFactory.create(line.split("\\^")));

        List<StructField> structFields = new ArrayList<StructField>();
        structFields.add(DataTypes.createStructField( "a", DataTypes.StringType, true ));
        structFields.add(DataTypes.createStructField( "b", DataTypes.StringType, true ));
        structFields.add(DataTypes.createStructField( "c", DataTypes.StringType, true ));
        StructType structType = DataTypes.createStructType( structFields );

        DataFrame test = sqlContext.createDataFrame( rows, structType);
        test.registerTempTable("test");

        sqlContext.sql("SELECT con_join(c,b) FROM test GROUP BY a").show();

        sc.stop();
    }

}

这样SQL简洁明了，就能表达意思了。

参考

• Spark Multiple Input/Output User Defined Aggregate Function (UDAF) using Java
• 李震的UDAF·scala版本
• Spark Sql官方文档
• Scala菜鸟教程
• spark1.5 自定义聚合函数UDAF