Spark SQL源码函数解读及UDF/UDAF例子 spark研习第六集

四、 Spark SQL源码函数解读

1. Spark SQL内置函数解密与实战

SparkSQL的DataFrame引入了大量的内置函数，这些内置函数一般都有CG（CodeGeneration）功能，这样的函数在编译和执行时都会经过高度优化。

问题：SparkSQL操作Hive和Hive on Spark一样吗？

=> 不一样。SparkSQL操作Hive只是把Hive当作数据仓库的来源，而计算引擎就是SparkSQL本身。Hive on spark是Hive的子项目，Hive on Spark的核心是把Hive的执行引擎换成Spark。众所周知，目前Hive的计算引擎是Mapreduce，因为性能低下等问题，所以Hive的官方就想替换这个引擎。

SparkSQL操作Hive上的数据叫Spark on Hive，而Hive on Spark依旧是以Hive为核心，只是把计算引擎由MapReduce替换为Spark。

Spark官网上DataFrame 的API Docs： 
http://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.package

Experimental 
A distributed collection of data organized into named columns. 
A DataFrame is equivalent to a relational table in Spark SQL. The following example creates a DataFrame by pointing Spark SQL to a Parquet data set.

 
 
val people = sqlContext.read.parquet("...")  // in Scala
DataFrame people = sqlContext.read().parquet("...")  // in Java
Once created, it can be manipulated using the various domain-specific-language (DSL) functions defined in: DataFrame (this class), Column, and functions.
To select a column from the data frame, use apply method in Scala and col in Java.
val ageCol = people("age")  // in Scala
Column ageCol = people.col("age")  // in Java
Note that the Column type can also be manipulated through its various functions.
// The following creates a new column that increases everybody's age by 10.
people("age") + 10  // in Scala
people.col("age").plus(10);  // in Java
A more concrete example in Scala:
// To create DataFrame using SQLContextval people = sqlContext.read.parquet("...")val department = sqlContext.read.parquet("...")
people.filter("age > 30")
  .join(department, people("deptId") === department("id"))
  .groupBy(department("name"), "gender")
  .agg(avg(people("salary")), max(people("age")))
and in Java:
// To create DataFrame using SQLContext
DataFrame people = sqlContext.read().parquet("...");
DataFrame department = sqlContext.read().parquet("...");
people.filter("age".gt(30))
  .join(department, people.col("deptId").equalTo(department("id")))
  .groupBy(department.col("name"), "gender")
  .agg(avg(people.col("salary")), max(people.col("age")));

以上内容中的join,groupBy,agg都是SparkSQL的内置函数。 
SParkl1.5.x以后推出了很多内置函数，据不完全统计，有一百多个内置函数。 
下面实战开发一个聚合操作的例子：

 
 
package com.dt.spark
import org.apache.spark.sql.types.{IntegerType, StringType, StructField, StructType}
import org.apache.spark.{SparkConf, SparkContext}
import org.apache.spark.sql.{Row, SQLContext}
import org.apache.spark.sql.functions._
/**
  * 使用Spark SQL中的内置函数对数据进行分析，Spark SQL API不同的是，DataFrame中的内置函数操作的结果是返回一个Column对象，而
  * DataFrame天生就是"A distributed collection of data organized into named columns.",这就为数据的复杂分析建立了坚实的基础
  * 并提供了极大的方便性，例如说，我们在操作DataFrame的方法中可以随时调用内置函数进行业务需要的处理，这之于我们构建附件的业务逻辑而言是可以
  * 极大的减少不必须的时间消耗（基于上就是实际模型的映射），让我们聚焦在数据分析上，这对于提高工程师的生产力而言是非常有价值的
  * Spark 1.5.x开始提供了大量的内置函数，例如agg：
  * def agg(aggExpr: (String, String), aggExprs: (String, String)*): DataFrame = {
  *  groupBy().agg(aggExpr, aggExprs : _*)
  *}
  * 还有max、mean、min、sum、avg、explode、size、sort_array、day、to_date、abs、acros、asin、atan
  * 总体上而言内置函数包含了五大基本类型：
  * 1，聚合函数，例如countDistinct、sumDistinct等；
  * 2，集合函数，例如sort_array、explode等
  * 3，日期、时间函数，例如hour、quarter、next_day
  * 4, 数学函数，例如asin、atan、sqrt、tan、round等；
  * 5，开窗函数，例如rowNumber等
  * 6，字符串函数，concat、format_number、rexexp_extract
  * 7, 其它函数，isNaN、sha、randn、callUDF
  */
object SparkSQLAgg {
  def main(args: Array[String]) {
    System.setProperty("hadoop.home.dir", "G:/datarguru spark/tool/hadoop-2.6.0")
    val conf = new SparkConf()
    conf.setAppName("SparkSQLlinnerFunctions")
    //conf.setMaster("spark://master:7077")
    conf.setMaster("local")
    val sc = new SparkContext(conf)
    val sqlContext = new SQLContext(sc) //构建SQL上下文
    //要使用Spark SQL的内置函数，就一定要导入SQLContext下的隐式转换
    import sqlContext.implicits._
    //模拟电商访问的数据，实际情况会比模拟数据复杂很多，最后生成RDD
    val userData = Array(
      "2016-3-27,001,http://spark.apache.org/,1000",
      "2016-3-27,001,http://Hadoop.apache.org/,1001",
      "2016-3-27,002,http://fink.apache.org/,1002",
      "2016-3-28,003,http://kafka.apache.org/,1020",
      "2016-3-28,004,http://spark.apache.org/,1010",
      "2016-3-28,002,http://hive.apache.org/,1200",
      "2016-3-28,001,http://parquet.apache.org/,1500",
      "2016-3-28,001,http://spark.apache.org/,1800"
    )
    val userDataRDD = sc.parallelize(userData)//生成分布式集群对象
    //根据业务需要对数据进行预处理生成DataFrame，要想把RDD转换成DataFrame，需要先把RDD中的元素类型变成Row类型
    //于此同时要提供DataFrame中的Columns的元数据信息描述
    val userDataRDDRow = userDataRDD.map(row => {val splited = row.split(","); Row(splited(0),splited(1).toInt,splited(2), splited(3).toInt)})
    val structType = StructType(Array(
      StructField("time", StringType, true),
      StructField("id", IntegerType, true),
      StructField("url", StringType, true),
      StructField("amount", IntegerType, true)
    ))
    val userDataDF = sqlContext.createDataFrame(userDataRDDRow, structType)
    //第五步：使用Spark SQL提供的内置函数对DataFrame进行操作，特别注意：内置函数生成的Column对象且自定进行CG；
    userDataDF.groupBy("time").agg('time, countDistinct('id))
        .map(row => Row(row(1),row(2))).collect().foreach(println)
    userDataDF.groupBy("time").agg('time, sum('amount))
      .map(row => Row(row(1),row(2))).collect().foreach(println)
  }
}

2. Spark SQL窗口函数解密与实战

窗口函数包括: 
分级函数、分析函数、聚合函数 
较全的窗口函数介绍参考： 
https://jaceklaskowski.gitbooks.io/mastering-apache-spark/content/spark-sql-windows.html

窗口函数中最重要的是row_number。row_bumber是对分组进行排序，所谓分组排序就是说在分组的基础上再进行排序。 
下面使用SparkSQL的方式重新编写TopNGroup.scala程序并执行：

 
 
package com.dt.spark
import org.apache.spark.sql.hive.HiveContext
import org.apache.spark.{SparkConf, SparkContext}
object SparkSQLWindowFunctionOps {
  def main(args: Array[String]) {
    val conf = new SparkConf()
    conf.setMaster("spark://master:7077")
    conf.setAppName("SparkSQLWindowFunctionOps")
    val sc = new SparkContext(conf)
    val hiveContext =  new HiveContext(sc)
    hiveContext.sql("DROP TABLE IF EXISTS scores")
    hiveContext.sql("CREATE TABLE IF NOT EXISTS  scores(name STRING,score INT)"
      +"ROW FORMAT DELIMITED FIELDS TERMINATED ' ' LINES TERMINATED BY '\\n'")
    //将要处理的数据导入到Hive表中
    hiveContext.sql("LOAD DATA LOCAL INPATH 'G://datarguru spark/tool/topNGroup.txt' INTO TABLE SCORES")
    //hiveContext.sql("LOAD DATA LOCAL INPATH '/opt/spark-1.4.0-bin-hadoop2.6/dataSource' INTO TABLE SCORES")
    /**
      * 使用子查询的方式完成目标数据的提取，在目标数据内幕使用窗口函数row_number来进行分组排序：
      * PARTITION BY :指定窗口函数分组的Key；
      * ORDER BY：分组后进行排序；
      */
    val result = hiveContext.sql("SELECT name,score FROM ("
      + "SELECT name,score,row_number() OVER (PARTITION BY name ORDER BY score DESC) rank FROM scores) sub_scores"
      + "WHERE rank <= 4")
    result.show() //在Driver的控制台上打印出结果内容
    //把数据保存在Hive数据仓库中
    hiveContext.sql("DROP TABLE IF EXISTS sortedResultScores")
    result.saveAsTable("sortedResultScores")
  }
}

报错：

 
 
ERROR metadata.Hive: NoSuchObjectException(message:default.scores table not found)
Exception in thread "main" org.apache.spark.sql.AnalysisException: missing BY at '' '' near '<EOF>'; line 1 pos 96

参考： 
http://blog.csdn.net/slq1023/article/details/51138709

3. Spark SQL UDF和UDAF解密与实战

UDAF=USER DEFINE AGGREGATE FUNCTION 
通过案例实战Spark SQL下的UDF和UDAF的具体使用： 
* UDF: User Defined Function，用户自定义的函数，函数的输入是一条具体的数据记录，实现上讲就是普通的Scala函数； 
* UDAF：User Defined Aggregation Function，用户自定义的聚合函数，函数本身作用于数据集合，能够在聚合操作的基础上进行自定义操作； 
* 实质上讲，例如说UDF会被Spark SQL中的Catalyst封装成为Expression，最终会通过eval方法来计算输入的数据Row（此处的Row和DataFrame中的Row没有任何关系）

1)实战编写UDF和UDAF：

 
 
package com.dt.spark
import org.apache.spark.sql.expressions.{MutableAggregationBuffer, UserDefinedAggregateFunction}
import org.apache.spark.sql.types._
import org.apache.spark.sql.{Row, SQLContext}
import org.apache.spark.{SparkConf, SparkContext}
object SparkSQLUDFUDAF {
  def main(args: Array[String]) {
    System.setProperty("hadoop.home.dir", "G:/datarguru spark/tool/hadoop-2.6.0");
    val conf = new SparkConf()
    conf.setAppName("SparkSQLUDFUDAF")
    conf.setMaster("local")
    val sc = new SparkContext(conf)
    val sqlContext = new SQLContext(sc)
    //模拟实际使用数据
    val bigData = Array("Spark", "Spark", "Hadoop", "Spark", "Hadoop", "Spark", "Spark", "Hadoop", "Spark", "Hadoop")
    //基于提供的数据创建DataFrame
    val bigDataRDD = sc.parallelize(bigData)
    val bigDataRow = bigDataRDD.map(item => Row(item))
    val structType = StructType(Array(StructField("word", StringType, true)))
    val bigDataDF = sqlContext.createDataFrame(bigDataRow, structType)
    bigDataDF.registerTempTable("bigDataTable") //注册成为临时表
    //通过SQLContext注册UDF，在Scala 2.10.x版本UDF函数最多可以接受22个输入参数
    sqlContext.udf.register("computeLength", (input: String) => input.length)
    //直接在SQL语句中使用UDF，就像使用SQL自动的内部函数一样
    sqlContext.sql("select word, computeLength(word) as length from bigDataTable").show()
    sqlContext.udf.register("wordCount", new MyUDAF)
    sqlContext.sql("select word,wordCount(word) as count,computeLength(word) " +
      "as length from bigDataTable group by word").show()
    while(true){}
  }
}
class MyUDAF extends UserDefinedAggregateFunction{ //ctrl+I实现复写方法
  /**
    * 该方法指定具体输入数据的类型
    * @return
    */
  override def inputSchema: StructType = StructType(Array(StructField("input", StringType, true)))
  /**
    * 在进行聚合操作的时候要处理的数据的结果的类型
    * @return
    */
  override def bufferSchema: StructType = StructType(Array(StructField("count", IntegerType, true)))
  /**
    * 指定UDAF函数计算后返回的结果类型
    * @return
    */
  override def dataType: DataType = IntegerType
  override def deterministic: Boolean = true
  /**
    * 在Aggregate之前每组数据的初始化结果
    * @param buffer
    * @param input
    */
  override def initialize(buffer: MutableAggregationBuffer): Unit = {buffer(0)=0}
  /**
    * 在进行聚合的时候有新的值进来，对分组后的聚合如何进行计算
    * 本地的聚合操作，相当于Hadoop MapReduce模型中的Combiner(这里的Row跟DataFrame的Row无关)
    * @param buffer
    * @param input
    */
  override def update(buffer: MutableAggregationBuffer, input: Row): Unit = {
    buffer(0) = buffer.getAs[Int](0) + 1
  }
  /**
    * 最后在分布式节点进行Local Reduce完成后需要进行全局级别的Merge操作
    * @param buffer1
    * @param buffer2
    */
  override def merge(buffer1: MutableAggregationBuffer, buffer2: Row): Unit = {
    buffer1(0) = buffer1.getAs[Int](0) + buffer2.getAs[Int](0)
  }
  /**
    *  返回UDAF最后的计算结果
    * @param buffer
    * @return
    */
  override def evaluate(buffer: Row): Any = buffer.getAs[Int](0)
}

2) UDFRegistration的源码：

 
 
/**
 * Functions for registering user-defined functions. Use [[SQLContext.udf]] to access this.
 *
 * @since 1.3.0
 */
class UDFRegistration private[sql] (sqlContext: SQLContext) extends Logging {
  private val functionRegistry = sqlContext.functionRegistry
  protected[sql] def registerPython(name: String, udf: UserDefinedPythonFunction): Unit = {
    log.debug(
      s"""
        | Registering new PythonUDF:
        | name: $name
        | command: ${udf.command.toSeq}
        | envVars: ${udf.envVars}
        | pythonIncludes: ${udf.pythonIncludes}
        | pythonExec: ${udf.pythonExec}
        | dataType: ${udf.dataType}
      """.stripMargin)
    functionRegistry.registerFunction(name, udf.builder)
  }
  /**
   * Register a user-defined aggregate function (UDAF).
   *
   * @param name the name of the UDAF.
   * @param udaf the UDAF needs to be registered.
   * @return the registered UDAF.
   */
  def register(
      name: String,
      udaf: UserDefinedAggregateFunction): UserDefinedAggregateFunction = {
    def builder(children: Seq[Expression]) = ScalaUDAF(children, udaf)
    functionRegistry.registerFunction(name, builder)
    udaf
  }
// scalastyle:off
/* register 0-22 were generated by this script
  (0 to 22).map { x =>
    val types = (1 to x).foldRight("RT")((i, s) => {s"A$i, $s"})
    val typeTags = (1 to x).map(i => s"A${i}: TypeTag").foldLeft("RT: TypeTag")(_ + ", " + _)
    val inputTypes = (1 to x).foldRight("Nil")((i, s) => {s"ScalaReflection.schemaFor[A$i].dataType :: $s"})
    println(s"""
      /**
       * Register a Scala closure of ${x} arguments as user-defined function (UDF).
       * @tparam RT return type of UDF.
       * @since 1.3.0
       */
      def register[$typeTags](name: String, func: Function$x[$types]): UserDefinedFunction = {
        val dataType = ScalaReflection.schemaFor[RT].dataType
        val inputTypes = Try($inputTypes).getOrElse(Nil)
        def builder(e: Seq[Expression]) = ScalaUDF(func, dataType, e, inputTypes)
        functionRegistry.registerFunction(name, builder)
        UserDefinedFunction(func, dataType, inputTypes)
      }""")
  }
  (1 to 22).foreach { i =>
    val extTypeArgs = (1 to i).map(_ => "_").mkString(", ")
    val anyTypeArgs = (1 to i).map(_ => "Any").mkString(", ")
    val anyCast = s".asInstanceOf[UDF$i[$anyTypeArgs, Any]]"
    val anyParams = (1 to i).map(_ => "_: Any").mkString(", ")
    println(s"""
       |/**
       | * Register a user-defined function with ${i} arguments.
       | * @since 1.3.0
       | */
       |def register(name: String, f: UDF$i[$extTypeArgs, _], returnType: DataType) = {
       |  functionRegistry.registerFunction(
       |    name,
       |    (e: Seq[Expression]) => ScalaUDF(f$anyCast.call($anyParams), returnType, e))
       |}""".stripMargin)
  }
  */
/**
 * Register a Scala closure of 0 arguments as user-defined function (UDF).
 * @tparam RT return type of UDF.
 * @since 1.3.0
 */
def register[RT: TypeTag](name: String, func: Function0[RT]): UserDefinedFunction = {
  val dataType = ScalaReflection.schemaFor[RT].dataType
  val inputTypes = Try(Nil).getOrElse(Nil)
  def builder(e: Seq[Expression]) = ScalaUDF(func, dataType, e, inputTypes)
  functionRegistry.registerFunction(name, builder)
  UserDefinedFunction(func, dataType, inputTypes)
}

FunctionRegistry的源码如下：

 
 
object FunctionRegistry {
  type FunctionBuilder = Seq[Expression] => Expression
  val expressions: Map[String, (ExpressionInfo, FunctionBuilder)] = Map(
    // misc non-aggregate functions
    expression[Abs]("abs"),
    expression[CreateArray]("array"),
    expression[Coalesce]("coalesce"),
    expression[Explode]("explode"),
    expression[Greatest]("greatest"),
    expression[If]("if"),
    expression[IsNaN]("isnan"),
    expression[IsNull]("isnull"),
    expression[IsNotNull]("isnotnull"),
    expression[Least]("least"),
    expression[Coalesce]("nvl"),
    expression[Rand]("rand"),
    expression[Randn]("randn"),
    expression[CreateStruct]("struct"),
    expression[CreateNamedStruct]("named_struct"),
    expression[Sqrt]("sqrt"),
    expression[NaNvl]("nanvl"),
    // math functions
    expression[Acos]("acos"),
    expression[Asin]("asin"),
    expression[Atan]("atan"),
    expression[Atan2]("atan2"),
    expression[Bin]("bin"),
    expression[Cbrt]("cbrt"),
    expression[Ceil]("ceil"),
    expression[Ceil]("ceiling"),
    expression[Cos]("cos"),
    expression[Cosh]("cosh"),
    expression[Conv]("conv"),
    expression[EulerNumber]("e"),
    expression[Exp]("exp"),
    expression[Expm1]("expm1"),
    expression[Floor]("floor"),
    expression[Factorial]("factorial"),
    expression[Hypot]("hypot"),
    expression[Hex]("hex"),
    expression[Logarithm]("log"),
    expression[Log]("ln"),
    expression[Log10]("log10"),
    expression[Log1p]("log1p"),
    expression[Log2]("log2"),
    expression[UnaryMinus]("negative"),
    expression[Pi]("pi"),
    expression[Pow]("pow"),
    expression[Pow]("power"),
    expression[Pmod]("pmod"),
    expression[UnaryPositive]("positive"),
    expression[Rint]("rint"),
    expression[Round]("round"),
    expression[ShiftLeft]("shiftleft"),
    expression[ShiftRight]("shiftright"),
    expression[ShiftRightUnsigned]("shiftrightunsigned"),
    expression[Signum]("sign"),
    expression[Signum]("signum"),
    expression[Sin]("sin"),
    expression[Sinh]("sinh"),
    expression[Tan]("tan"),
    expression[Tanh]("tanh"),
    expression[ToDegrees]("degrees"),
    expression[ToRadians]("radians"),
    // aggregate functions
    expression[HyperLogLogPlusPlus]("approx_count_distinct"),
    expression[Average]("avg"),
    expression[Corr]("corr"),
    expression[Count]("count"),
    expression[First]("first"),
    expression[First]("first_value"),
    expression[Last]("last"),
    expression[Last]("last_value"),
    expression[Max]("max"),
    expression[Average]("mean"),
    expression[Min]("min"),
    expression[StddevSamp]("stddev"),
    expression[StddevPop]("stddev_pop"),
    expression[StddevSamp]("stddev_samp"),
    expression[Sum]("sum"),
    expression[VarianceSamp]("variance"),
    expression[VariancePop]("var_pop"),
    expression[VarianceSamp]("var_samp"),
    expression[Skewness]("skewness"),
    expression[Kurtosis]("kurtosis"),
    // string functions
    expression[Ascii]("ascii"),
    expression[Base64]("base64"),
    expression[Concat]("concat"),
    expression[ConcatWs]("concat_ws"),
    expression[Encode]("encode"),
    expression[Decode]("decode"),
    expression[FindInSet]("find_in_set"),
    expression[FormatNumber]("format_number"),
    expression[GetJsonObject]("get_json_object"),
    expression[InitCap]("initcap"),
    expression[JsonTuple]("json_tuple"),
    expression[Lower]("lcase"),
    expression[Lower]("lower"),
    expression[Length]("length"),
    expression[Levenshtein]("levenshtein"),
    expression[RegExpExtract]("regexp_extract"),
    expression[RegExpReplace]("regexp_replace"),
    expression[StringInstr]("instr"),
    expression[StringLocate]("locate"),
    expression[StringLPad]("lpad"),
    expression[StringTrimLeft]("ltrim"),
    expression[FormatString]("format_string"),
    expression[FormatString]("printf"),
    expression[StringRPad]("rpad"),
    expression[StringRepeat]("repeat"),
    expression[StringReverse]("reverse"),
    expression[StringTrimRight]("rtrim"),
    expression[SoundEx]("soundex"),
    expression[StringSpace]("space"),
    expression[StringSplit]("split"),
    expression[Substring]("substr"),
    expression[Substring]("substring"),
    expression[SubstringIndex]("substring_index"),
    expression[StringTranslate]("translate"),
    expression[StringTrim]("trim"),
    expression[UnBase64]("unbase64"),
    expression[Upper]("ucase"),
    expression[Unhex]("unhex"),
    expression[Upper]("upper"),
...

可以看出SparkSQL的内置函数也是和UDF一样注册的。

4. Spark SQL Thrift Server实战

The Thrift JDBC/ODBC server implemented here corresponds to the HiveServer2 in Hive 1.2.1 You can test the JDBC server with the beeline script that comes with either Spark or Hive 1.2.1. 
打开JDBC/ODBC server：

 
 
ps -aux | grep hive 
 hive --service metastore &  //先打开hive元数据
[1] 28268
./sbin/start-thriftserver.sh
//Now you can use beeline to test the Thrift JDBC/ODBC server:
./bin/beeline
//Connect to the JDBC/ODBC server in beeline with:
beeline> !connect jdbc:hive2://master:10000
//:root
//密码为空
hive命令

Java通过JDBC访问Thrift Server

 
 
package com.dt.sparksql;
import java.sql.Connection;
import java.sql.DriverManager;
import java.sql.PreparedStatement;
import java.sql.ResultSet;
import java.sql.SQLException;
/**
 * 演示Java通过JDBC访问Thrift  Server，进而访问Spark SQL，进而访问Hive，这是企业级开发中最为常见的方式
 * @author dt_sparl
 *
 */
public class SparkSQLJDBC2ThriftServer {
    public static void main(String[] args) throws SQLException {
        String sqlTest = "select name from people where age = ?";
        Connection conn = null;
        ResultSet resultSet = null;
        try {
            Class.forName("org.apache.hive.jdbc.HiveDriver");
            conn = DriverManager.getConnection("jdbc:hive2://<master>:<10001>/<default>?"
                + "hive.server2.transport.mode=http;hive.server2.thrift.http.path=<cliserver>",
                "root", "");
             PreparedStatement  preparedStatement = conn.prepareStatement(sqlTest);
             preparedStatement.setInt(1, 30);
             resultSet = preparedStatement.executeQuery();
             while(resultSet.next()){
                 System.out.println(resultSet.getString(1)); //这里的数据应该保存在parquet中
             }
        } catch (ClassNotFoundException e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        }finally {
            resultSet.close();
            conn.close();
        }
    }
}