Hive-UDAF

UDAF

前两节分别介绍了基础UDF和UDTF，这一节我们将介绍最复杂的用户自定义聚合函数（UDAF）。用户自定义聚合函数（UDAF）接受从零行到多行的零个到多个列，然后返回单一值，如sum()、count()。要实现UDAF，我们需要实现下面的类：

org.apache.hadoop.hive.ql.udf.generic.AbstractGenericUDAFResolver

org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator

AbstractGenericUDAFResolver检查输入参数，并且指定使用哪个resolver。在AbstractGenericUDAFResolver里，只需要实现一个方法：

public GenericUDAFEvaluator getEvaluator(TypeInfo[] parameters) throws SemanticException;  

但是，主要的逻辑处理还是在Evaluator中。我们需要继承GenericUDAFEvaluator，并且实现下面几个方法：

// 输入输出都是Object inspectors  
public  ObjectInspector init(Mode m, ObjectInspector[] parameters) throws HiveException;  
  
// AggregationBuffer保存数据处理的临时结果  
abstract AggregationBuffer getNewAggregationBuffer() throws HiveException;  
  
// 重新设置AggregationBuffer  
public void reset(AggregationBuffer agg) throws HiveException;  
  
// 处理输入记录  
public void iterate(AggregationBuffer agg, Object[] parameters) throws HiveException;  
  
// 处理全部输出数据中的部分数据  
public Object terminatePartial(AggregationBuffer agg) throws HiveException;  
  
// 把两个部分数据聚合起来  
public void merge(AggregationBuffer agg, Object partial) throws HiveException;  
  
// 输出最终结果  
public Object terminate(AggregationBuffer agg) throws HiveException;  

在处理之前，先看下UADF的Enum GenericUDAFEvaluator.Mode。Mode有4中情况：

• PARTIAL1：Mapper阶段。从原始数据到部分聚合，会调用iterate()和terminatePartial()。
• PARTIAL2：Combiner阶段，在Mapper端合并Mapper的结果数据。从部分聚合到部分聚合，会调用merge()和terminatePartial()。
• FINAL：Reducer阶段。从部分聚合数据到完全聚合，会调用merge()和terminate()。
• COMPLETE：出现这个阶段，表示MapReduce中只用Mapper没有Reducer，所以Mapper端直接输出结果了。从原始数据到完全聚合，会调用iterate()和terminate()。

GenericUDAFResolver2

@Deprecated
public abstract interface GenericUDAFResolver {
    public abstract GenericUDAFEvaluator getEvaluator(TypeInfo[] paramArrayOfTypeInfo) throws SemanticException;
}

已废弃

public abstract interface GenericUDAFResolver2 extends GenericUDAFResolver {
    public abstract GenericUDAFEvaluator getEvaluator(GenericUDAFParameterInfo paramGenericUDAFParameterInfo)
            throws SemanticException;
}

GenericUDAFEvaluator

@UDFType(deterministic = true)
public abstract class GenericUDAFEvaluator implements Closeable {
    Mode mode;

    public static boolean isEstimable(AggregationBuffer buffer) {
        if (buffer instanceof AbstractAggregationBuffer) {
            Class clazz = buffer.getClass();
            AggregationType annotation = (AggregationType) clazz.getAnnotation(AggregationType.class);
            return ((annotation != null) && (annotation.estimable()));
        }
        return false;
    }

    public void configure(MapredContext mapredContext) {
    }

    public ObjectInspector init(Mode m, ObjectInspector[] parameters) throws HiveException {
        this.mode = m;
        return null;
    }

    public abstract AggregationBuffer getNewAggregationBuffer() throws HiveException;

    public abstract void reset(AggregationBuffer paramAggregationBuffer) throws HiveException;

    public void close() throws IOException {
    }

    public void aggregate(AggregationBuffer agg, Object[] parameters) throws HiveException {
        if ((this.mode == Mode.PARTIAL1) || (this.mode == Mode.COMPLETE)) {
            iterate(agg, parameters);
        } else {
            assert (parameters.length == 1);
            merge(agg, parameters[0]);
        }
    }

    public Object evaluate(AggregationBuffer agg) throws HiveException {
        if ((this.mode == Mode.PARTIAL1) || (this.mode == Mode.PARTIAL2)) {
            return terminatePartial(agg);
        }
        return terminate(agg);
    }

    public abstract void iterate(AggregationBuffer paramAggregationBuffer, Object[] paramArrayOfObject)
            throws HiveException;

    public abstract Object terminatePartial(AggregationBuffer paramAggregationBuffer) throws HiveException;

    public abstract void merge(AggregationBuffer paramAggregationBuffer, Object paramObject) throws HiveException;

    public abstract Object terminate(AggregationBuffer paramAggregationBuffer) throws HiveException;

    public static abstract class AbstractAggregationBuffer implements GenericUDAFEvaluator.AggregationBuffer {
        public int estimate() {
            return -1;
        }
    }

    public static abstract interface AggregationBuffer {
    }

    public static enum Mode {
        PARTIAL1, PARTIAL2, FINAL, COMPLETE;
    }

    public static @interface AggregationType {
        public abstract boolean estimable();
    }
}

例子

count

/*** Eclipse Class Decompiler plugin, copyright (c) 2016 Chen Chao (cnfree2000@hotmail.com) ***/
package org.apache.hadoop.hive.ql.udf.generic;

import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.hive.ql.exec.Description;
import org.apache.hadoop.hive.ql.exec.UDFArgumentException;
import org.apache.hadoop.hive.ql.metadata.HiveException;
import org.apache.hadoop.hive.ql.parse.SemanticException;
import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.LongObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.PrimitiveObjectInspectorFactory;
import org.apache.hadoop.hive.serde2.typeinfo.TypeInfo;
import org.apache.hadoop.io.LongWritable;

@Description(name = "count", value = "_FUNC_(*) - Returns the total number of retrieved rows, including rows containing NULL values.\n_FUNC_(expr) - Returns the number of rows for which the supplied expression is non-NULL.\n_FUNC_(DISTINCT expr[, expr...]) - Returns the number of rows for which the supplied expression(s) are unique and non-NULL.")
public class GenericUDAFCount implements GenericUDAFResolver2 {
    private static final Log LOG;

    public GenericUDAFEvaluator getEvaluator(TypeInfo[] parameters) throws SemanticException {
        return new GenericUDAFCountEvaluator();
    }

    public GenericUDAFEvaluator getEvaluator(GenericUDAFParameterInfo paramInfo) throws SemanticException {
        TypeInfo[] parameters = paramInfo.getParameters();

        if (parameters.length == 0) {
            if (!(paramInfo.isAllColumns())) {
                throw new UDFArgumentException("Argument expected");
            }
            if ((!($assertionsDisabled)) && (paramInfo.isDistinct()))
                throw new AssertionError("DISTINCT not supported with *");
        } else {
            if ((parameters.length > 1) && (!(paramInfo.isDistinct()))) {
                throw new UDFArgumentException("DISTINCT keyword must be specified");
            }
            assert (!(paramInfo.isAllColumns())) : "* not supported in expression list";
        }

        return new GenericUDAFCountEvaluator().setCountAllColumns(paramInfo.isAllColumns());
    }

    static {
        LOG = LogFactory.getLog(GenericUDAFCount.class.getName());
    }

    public static class GenericUDAFCountEvaluator extends GenericUDAFEvaluator {
        private boolean countAllColumns;
        private LongObjectInspector partialCountAggOI;
        private LongWritable result;

        public GenericUDAFCountEvaluator() {
            this.countAllColumns = false;
        }

        public ObjectInspector init(GenericUDAFEvaluator.Mode m, ObjectInspector[] parameters) throws HiveException {
            super.init(m, parameters);
            this.partialCountAggOI = PrimitiveObjectInspectorFactory.writableLongObjectInspector;

            this.result = new LongWritable(0L);
            return PrimitiveObjectInspectorFactory.writableLongObjectInspector;
        }

        private GenericUDAFCountEvaluator setCountAllColumns(boolean countAllCols) {
            this.countAllColumns = countAllCols;
            return this;
        }

        public GenericUDAFEvaluator.AggregationBuffer getNewAggregationBuffer() throws HiveException {
            CountAgg buffer = new CountAgg();
            reset(buffer);
            return buffer;
        }

        public void reset(GenericUDAFEvaluator.AggregationBuffer agg) throws HiveException {
            ((CountAgg) agg).value = 0L;
        }

        public void iterate(GenericUDAFEvaluator.AggregationBuffer agg, Object[] parameters) throws HiveException {
            if (parameters == null) {
                return;
            }
            if (this.countAllColumns) {
                assert (parameters.length == 0);
                ((CountAgg) agg).value += 1L;
            } else {
                assert (parameters.length > 0);
                boolean countThisRow = true;
                for (Object nextParam : parameters) {
                    if (nextParam == null) {
                        countThisRow = false;
                        break;
                    }
                }
                if (countThisRow)
                    ((CountAgg) agg).value += 1L;
            }
        }

        public void merge(GenericUDAFEvaluator.AggregationBuffer agg, Object partial) throws HiveException {
            if (partial != null) {
                long p = this.partialCountAggOI.get(partial);
                ((CountAgg) agg).value += p;
            }
        }

        public Object terminate(GenericUDAFEvaluator.AggregationBuffer agg) throws HiveException {
            this.result.set(((CountAgg) agg).value);
            return this.result;
        }

        public Object terminatePartial(GenericUDAFEvaluator.AggregationBuffer agg) throws HiveException {
            return terminate(agg);
        }

        @GenericUDAFEvaluator.AggregationType(estimable = true)
        static class CountAgg extends GenericUDAFEvaluator.AbstractAggregationBuffer {
            long value;

            public int estimate() {
                return 8;
            }
        }
    }
}

sum

udaf 需要hive的sql和group by联合使用。hive的group by对于每个分组，只能返回一条记录。

开发通用udaf有另个步骤，一个是编写resolver类，第二个是编写evaluator类。resolver负责类型检查，操作符重载。evaluator负责实现真正的udaf逻辑、

以sum为例、

reslver通常继承resolver2.但是建议继承resolver。隔离将来hive接口的变化。

public class GenericUDAFSum extends AbstractGenericUDAFResolver {
    static final Log LOG = LogFactory.getLog(GenericUDAFSum.class.getName());

    public GenericUDAFEvaluator getEvaluator(TypeInfo[] parameters)
    throws SemanticException
  {
    if (parameters.length != 1) {
      throw new UDFArgumentTypeException(parameters.length - 1, "Exactly one argument is expected.");
    }

    if (parameters[0].getCategory() != ObjectInspector.Category.PRIMITIVE) {
      throw new UDFArgumentTypeException(0, "Only primitive type arguments are accepted but " + parameters[0].getTypeName() + " is passed.");
    }

    switch (1.$SwitchMap$org$apache$hadoop$hive$serde2$objectinspector$PrimitiveObjectInspector$PrimitiveCategory[((org.apache.hadoop.hive.serde2.typeinfo.PrimitiveTypeInfo)parameters[0]).getPrimitiveCategory().ordinal()]) {
    case 1:
    case 2:
    case 3:
    case 4:
      return new GenericUDAFSumLong();
    case 5:
    case 6:
    case 7:
    case 8:
    case 9:
    case 10:
      return new GenericUDAFSumDouble();
    case 11:
      return new GenericUDAFSumHiveDecimal();
    case 12:
    case 13:
    }
    throw new UDFArgumentTypeException(0, "Only numeric or string type arguments are accepted but " + parameters[0].getTypeName() + " is passed.");
  }

着就是udaf的代码骨架。创建一个log对象。 重写getEvaluator方法。根据sql传入的参数类型，返回争取的evaluator。主要实现操作符的重载。

实现evaluator

下面以genericudafsumlong为例。

public static class GenericUDAFSumLong extends GenericUDAFEvaluator {
        private PrimitiveObjectInspector inputOI;
        private LongWritable result;
        private boolean warned;

        public GenericUDAFSumLong() {
            this.warned = false;
        }
        //这个方法返回可udaf的返回类型。这里定义返回类型为long
        public ObjectInspector init(GenericUDAFEvaluator.Mode m, ObjectInspector[] parameters) throws HiveException {
            assert (parameters.length == 1);
            super.init(m, parameters);
            this.result = new LongWritable(0L);
            this.inputOI = ((PrimitiveObjectInspector) parameters[0]);
            return PrimitiveObjectInspectorFactory.writableLongObjectInspector;
        }

        //创建新的聚合计算需要的内存，用来存储mapper，combiner，reducer运算过程中的相加总和。
        public GenericUDAFEvaluator.AggregationBuffer getNewAggregationBuffer() throws HiveException {
            SumLongAgg result = new SumLongAgg();
            reset(result);
            return result;
        }

        //mr支持mapper和reducer的重用，所以为了兼容，也要做内存的重用
        public void reset(GenericUDAFEvaluator.AggregationBuffer agg) throws HiveException {
            SumLongAgg myagg = (SumLongAgg) agg;
            myagg.empty = true;
            myagg.sum = 0L;
        }
        
        //map阶段，只要把保存道歉和的对象agg，再加上输入的参数，就可以了。
        public void iterate(GenericUDAFEvaluator.AggregationBuffer agg, Object[] parameters) throws HiveException {
            assert (parameters.length == 1);
            try {
                merge(agg, parameters[0]);
            } catch (NumberFormatException e) {
                if (!(this.warned)) {
                    this.warned = true;
                    GenericUDAFSum.LOG.warn(super.getClass().getSimpleName() + " " + StringUtils.stringifyException(e));
                }
            }
        }

        //mapper结束要返回的结果和combiner结束要返回的结果。
        public Object terminatePartial(GenericUDAFEvaluator.AggregationBuffer agg) throws HiveException {
            return terminate(agg);
        }
        
        //combiner合并map返回的结果，还有reducer合并mapper或combiner返回的结果
        public void merge(GenericUDAFEvaluator.AggregationBuffer agg, Object partial) throws HiveException {
            if (partial != null) {
                SumLongAgg myagg = (SumLongAgg) agg;
                myagg.sum += PrimitiveObjectInspectorUtils.getLong(partial, this.inputOI);
                myagg.empty = false;
            }
        }

        //reducer返回结果，或者是只有mapper，没有reducer，在mapper端返回结果。
        public Object terminate(GenericUDAFEvaluator.AggregationBuffer agg) throws HiveException {
            SumLongAgg myagg = (SumLongAgg) agg;
            if (myagg.empty) {
                return null;
            }
            this.result.set(myagg.sum);
            return this.result;
        }
        
        //存储sum值得类
        @GenericUDAFEvaluator.AggregationType(estimable = true)
        static class SumLongAgg extends GenericUDAFEvaluator.AbstractAggregationBuffer {
            boolean empty;
            long sum;

            public int estimate() {
                return 12;
            }
        }
    }