spark SQL常用聚合函数
This commit is contained in:
罗祥
parent
656090f384
commit
0dd9d8862f
14
README.md
14
README.md
@ -74,20 +74,22 @@
|
||||
|
||||
1. [Spark简介](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark简介.md)
|
||||
2. [Spark开发环境搭建](https://github.com/heibaiying/BigData-Notes/blob/master/notes/installation/Spark开发环境搭建.md)
|
||||
4. [弹性式数据集RDD](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark-RDD.md)
|
||||
5. [RDD常用算子详解](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark-Transformation和Action.md)
|
||||
4. [弹性式数据集RDD](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark_RDD.md)
|
||||
5. [RDD常用算子详解](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark_Transformation和Action算子.md)
|
||||
5. [Spark运行模式与作业提交](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark部署模式与作业提交.md)
|
||||
6. [Spark累加器与广播变量](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark累加器与广播变量.md)
|
||||
|
||||
**Spark SQL :**
|
||||
|
||||
1. [Spark SQL之 DateFrame 和 DataSet](https://github.com/heibaiying/BigData-Notes/blob/master/notes/SparkSQL-Dataset&DataFrame.md)
|
||||
2. Spark SQL之常用SQL语句
|
||||
3. External Data Source
|
||||
1. [DateFrames 和 DataSets ](https://github.com/heibaiying/BigData-Notes/blob/master/notes/SparkSQL_Dataset和DataFrame简介.md)
|
||||
2. [Structured API的基本使用](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Structured_API的基本使用.md)
|
||||
3. 外部数据源
|
||||
4. [Spark SQL常用聚合函数](https://github.com/heibaiying/BigData-Notes/blob/master/notes/SparkSQL常用聚合函数.md)
|
||||
5. 联结操作
|
||||
|
||||
**Spark Streaming ：**
|
||||
|
||||
1. [Spark Streaming简介](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark-Streaming与流处理.md)
|
||||
1. [Spark Streaming简介](https://github.com/heibaiying/BigData-Notes/blob/master/notes/Spark_Streaming与流处理.md)
|
||||
2. DStream常用操作详解
|
||||
3. Spark Streaming 整合 Flume
|
||||
4. Spark Streaming 整合 Kafka
|
||||
|
||||
@ -1,65 +1,54 @@
|
||||
package rdd.scala
|
||||
|
||||
import org.apache.spark.sql.expressions.Aggregator
|
||||
import org.apache.spark.sql.{Encoder, Encoders, SparkSession, functions}
|
||||
import org.apache.spark.sql.SparkSession
|
||||
import org.apache.spark.sql.functions._
|
||||
|
||||
// 1.定义员工类,对于可能存在null值的字段需要使用Option进行包装
|
||||
case class Emp(ename: String, comm: scala.Option[Double], deptno: Long, empno: Long,
|
||||
hiredate: String, job: String, mgr: scala.Option[Long], sal: Double)
|
||||
|
||||
// 2.定义聚合操作的中间输出类型
|
||||
case class SumAndCount(var sum: Double, var count: Long)
|
||||
|
||||
/* 3.自定义聚合函数
|
||||
* @IN 聚合操作的输入类型
|
||||
* @BUF reduction操作输出值的类型
|
||||
* @OUT 聚合操作的输出类型
|
||||
*/
|
||||
object MyAverage extends Aggregator[Emp, SumAndCount, Double] {
|
||||
|
||||
|
||||
// 4.用于聚合操作的的初始零值
|
||||
override def zero: SumAndCount = SumAndCount(0, 0)
|
||||
|
||||
|
||||
// 5.同一分区中的reduce操作
|
||||
override def reduce(avg: SumAndCount, emp: Emp): SumAndCount = {
|
||||
avg.sum += emp.sal
|
||||
avg.count += 1
|
||||
avg
|
||||
}
|
||||
|
||||
// 6.不同分区中的merge操作
|
||||
override def merge(avg1: SumAndCount, avg2: SumAndCount): SumAndCount = {
|
||||
avg1.sum += avg2.sum
|
||||
avg1.count += avg2.count
|
||||
avg1
|
||||
}
|
||||
|
||||
// 7.定义最终的输出类型
|
||||
override def finish(reduction: SumAndCount): Double = reduction.sum / reduction.count
|
||||
|
||||
// 8.中间类型的编码转换
|
||||
override def bufferEncoder: Encoder[SumAndCount] = Encoders.product
|
||||
|
||||
// 9.输出类型的编码转换
|
||||
override def outputEncoder: Encoder[Double] = Encoders.scalaDouble
|
||||
}
|
||||
|
||||
object SparkSqlApp {
|
||||
|
||||
// 测试方法
|
||||
def main(args: Array[String]): Unit = {
|
||||
|
||||
val spark = SparkSession.builder().appName("Spark-SQL").master("local[2]").getOrCreate()
|
||||
import spark.implicits._
|
||||
val ds = spark.read.json("file/emp.json").as[Emp]
|
||||
val spark = SparkSession.builder().appName("aggregations").master("local[2]").getOrCreate()
|
||||
val empDF = spark.read.json("/usr/file/json/emp.json")
|
||||
empDF.createOrReplaceTempView("emp")
|
||||
empDF.show()
|
||||
|
||||
empDF.select(count("ename")).show()
|
||||
empDF.select(countDistinct("deptno")).show()
|
||||
empDF.select(approx_count_distinct("ename", 0.1)).show()
|
||||
empDF.select(first("ename"), last("job")).show()
|
||||
empDF.select(min("sal"), max("sal")).show()
|
||||
empDF.select(sum("sal")).show()
|
||||
empDF.select(sumDistinct("sal")).show()
|
||||
empDF.select(avg("sal")).show()
|
||||
|
||||
|
||||
// 总体方差 均方差 总体标准差 样本标准差
|
||||
empDF.select(var_pop("sal"), var_samp("sal"), stddev_pop("sal"), stddev_samp("sal")).show()
|
||||
|
||||
|
||||
// 偏度和峰度
|
||||
empDF.select(skewness("sal"), kurtosis("sal")).show()
|
||||
|
||||
// 计算两列的 皮尔逊相关系数 样本协方差 总体协方差
|
||||
empDF.select(corr("empno", "sal"), covar_samp("empno", "sal"),
|
||||
covar_pop("empno", "sal")).show()
|
||||
|
||||
empDF.agg(collect_set("job"), collect_list("ename")).show()
|
||||
|
||||
|
||||
empDF.groupBy("deptno", "job").count().show()
|
||||
spark.sql("SELECT deptno, job, count(*) FROM emp GROUP BY deptno, job").show()
|
||||
|
||||
empDF.groupBy("deptno").agg(count("ename").alias("人数"), sum("sal").alias("总工资")).show()
|
||||
spark.sql("SELECT deptno, count(ename) ,sum(sal) FROM emp GROUP BY deptno").show()
|
||||
|
||||
|
||||
empDF.groupBy("deptno").agg("ename"->"count","sal"->"sum").show()
|
||||
|
||||
|
||||
|
||||
// 10.使用内置avg()函数和自定义函数分别进行计算，验证自定义函数是否正确
|
||||
val myAvg = ds.select(MyAverage.toColumn.name("average_sal")).first()
|
||||
val avg = ds.select(functions.avg(ds.col("sal"))).first().get(0)
|
||||
|
||||
println("自定义average函数 : " + myAvg)
|
||||
println("内置的average函数 : " + avg)
|
||||
}
|
||||
}
|
||||
|
||||
@ -1,345 +0,0 @@
|
||||
# SparkSQL API基本使用
|
||||
|
||||
## 一、创建DataFrames
|
||||
|
||||
Spark中所有功能的入口点是`SparkSession`，可以使用`SparkSession.builder()`创建。创建后应用程序就可以从现有RDD，Hive表或Spark数据源创建DataFrame。如下所示：
|
||||
|
||||
```scala
|
||||
val spark = SparkSession.builder().appName("Spark-SQL").master("local[2]").getOrCreate()
|
||||
val df = spark.read.json("/usr/file/emp.json")
|
||||
df.show()
|
||||
|
||||
// 建议在进行spark SQL编程前导入下面的隐式转换，因为DataFrames和dataSets中很多操作都依赖了隐式转换
|
||||
import spark.implicits._
|
||||
```
|
||||
|
||||
这里可以启动`spark-shell`进行测试，需要注意的是`spark-shell`启动后会自动创建一个名为`spark`的`SparkSession`，在命令行中可以直接引用即可：
|
||||
|
||||
|
||||
|
||||
## 二、DataFrames基本操作
|
||||
|
||||
### 2.1 printSchema
|
||||
|
||||
```scala
|
||||
// 以树形结构打印dataframe的schema信息
|
||||
df.printSchema()
|
||||
```
|
||||
|
||||
|
||||
|
||||
### 2.2 使用DataFrame API进行基本查询
|
||||
|
||||
```scala
|
||||
// 查询员工姓名及工作
|
||||
df.select($"ename", $"job").show()
|
||||
|
||||
// 查询工资大于2000的员工信息
|
||||
df.filter($"sal" > 2000).show()
|
||||
|
||||
// 分组统计部门人数
|
||||
df.groupBy("deptno").count().show()
|
||||
```
|
||||
|
||||
### 2.3 使用SQL进行基本查询
|
||||
|
||||
```scala
|
||||
// 首先需要将DataFrame注册为临时视图
|
||||
df.createOrReplaceTempView("emp")
|
||||
|
||||
// 查询员工姓名及工作
|
||||
spark.sql("SELECT ename,job FROM emp").show()
|
||||
|
||||
// 查询工资大于2000的员工信息
|
||||
spark.sql("SELECT * FROM emp where sal > 2000").show()
|
||||
|
||||
// 分组统计部门人数
|
||||
spark.sql("SELECT deptno,count(ename) FROM emp group by deptno").show()
|
||||
```
|
||||
|
||||
### 2.4 全局临时视图
|
||||
|
||||
上面使用`createOrReplaceTempView`创建的是会话临时视图，它的生命周期仅限于会话范围，会随会话的结束而结束。
|
||||
|
||||
你也可以使用`createGlobalTempView`创建全局临时视图，全局临时视图可以在所有会话之间共享，并直到整个Spark应用程序终止才会消失。全局临时视图被定义在内置的`global_temp`数据库下，需要使用限定名称进行引用，如`SELECT * FROM global_temp.view1`。
|
||||
|
||||
```scala
|
||||
// 注册为全局临时视图
|
||||
df.createGlobalTempView("gemp")
|
||||
|
||||
// 查询员工姓名及工作，使用限定名称进行引用
|
||||
spark.sql("SELECT ename,job FROM global_temp.gemp").show()
|
||||
|
||||
// 查询工资大于2000的员工信息，使用限定名称进行引用
|
||||
spark.sql("SELECT * FROM global_temp.gemp where sal > 2000").show()
|
||||
|
||||
// 分组统计部门人数，使用限定名称进行引用
|
||||
spark.sql("SELECT deptno,count(ename) FROM global_temp.gemp group by deptno").show()
|
||||
```
|
||||
|
||||
## 三、创建Datasets
|
||||
|
||||
### 3.1 由外部数据集创建
|
||||
|
||||
```scala
|
||||
// 1.需要导入隐式转换
|
||||
import spark.implicits._
|
||||
|
||||
// 2.创建case class,等价于Java Bean
|
||||
case class Emp(ename: String, comm: Double, deptno: Long, empno: Long,
|
||||
hiredate: String, job: String, mgr: Long, sal: Double)
|
||||
|
||||
// 3.由外部数据集创建Datasets
|
||||
val ds = spark.read.json("/usr/file/emp.json").as[Emp]
|
||||
ds.show()
|
||||
```
|
||||
|
||||
### 3.2 由内部数据集创建
|
||||
|
||||
```scala
|
||||
// 1.需要导入隐式转换
|
||||
import spark.implicits._
|
||||
|
||||
// 2.创建case class,等价于Java Bean
|
||||
case class Emp(ename: String, comm: Double, deptno: Long, empno: Long,
|
||||
hiredate: String, job: String, mgr: Long, sal: Double)
|
||||
|
||||
// 3.由内部数据集创建Datasets
|
||||
val caseClassDS = Seq(Emp("ALLEN", 300.0, 30, 7499, "1981-02-20 00:00:00", "SALESMAN", 7698, 1600.0),
|
||||
Emp("JONES", 300.0, 30, 7499, "1981-02-20 00:00:00", "SALESMAN", 7698, 1600.0))
|
||||
.toDS()
|
||||
caseClassDS.show()
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 四、DataFrames与Datasets互相转换
|
||||
|
||||
Spark提供了非常简单的转换方法用于DataFrames与Datasets互相转换，示例如下：
|
||||
|
||||
```shell
|
||||
# DataFrames转Datasets
|
||||
scala> df.as[Emp]
|
||||
res1: org.apache.spark.sql.Dataset[Emp] = [COMM: double, DEPTNO: bigint ... 6 more fields]
|
||||
|
||||
# Datasets转DataFrames
|
||||
scala> ds.toDF()
|
||||
res2: org.apache.spark.sql.DataFrame = [COMM: double, DEPTNO: bigint ... 6 more fields]
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 五、RDDs转换为DataFrames\Datasets
|
||||
|
||||
Spark支持两种方式把RDD转换为DataFrames，分别是使用反射推断和指定schema转换。
|
||||
|
||||
### 5.1 使用反射推断
|
||||
|
||||
```scala
|
||||
// 1.导入隐式转换
|
||||
import spark.implicits._
|
||||
|
||||
// 2.创建部门类
|
||||
case class Dept(deptno: Long, dname: String, loc: String)
|
||||
|
||||
// 3.创建RDD并转换为dataSet
|
||||
val rddToDS = spark.sparkContext
|
||||
.textFile("/usr/file/dept.txt")
|
||||
.map(_.split("\t"))
|
||||
.map(line => Dept(line(0).trim.toLong, line(1), line(2)))
|
||||
.toDS() // 如果调用toDF()则转换为dataFrame
|
||||
```
|
||||
|
||||
### 5.2 以编程方式指定Schema
|
||||
|
||||
```scala
|
||||
import org.apache.spark.sql.Row
|
||||
import org.apache.spark.sql.types._
|
||||
|
||||
|
||||
// 1.定义每个列的列类型
|
||||
val fields = Array(StructField("deptno", LongType, nullable = true),
|
||||
StructField("dname", StringType, nullable = true),
|
||||
StructField("loc", StringType, nullable = true))
|
||||
|
||||
// 2.创建schema
|
||||
val schema = StructType(fields)
|
||||
|
||||
// 3.创建RDD
|
||||
val deptRDD = spark.sparkContext.textFile("/usr/file/dept.txt")
|
||||
val rowRDD = deptRDD.map(_.split("\t")).map(line => Row(line(0).toLong, line(1), line(2)))
|
||||
|
||||
|
||||
// 4.将RDD转换为dataFrame
|
||||
val deptDF = spark.createDataFrame(rowRDD, schema)
|
||||
deptDF.show()
|
||||
```
|
||||
|
||||
## 六、使用自定义聚合函数
|
||||
|
||||
Scala提供了两种自定义聚合函数的方法，分别如下：
|
||||
|
||||
+ 有类型的自定义聚合函数，主要适用于DataSets；
|
||||
+ 无类型的自定义聚合函数，主要适用于DataFrames。
|
||||
|
||||
以下分别使用两种方式来自定义一个求平均值的聚合函数，这里以计算员工平均工资为例。两种自定义方式分别如下：
|
||||
|
||||
### 6.1 有类型的自定义函数
|
||||
|
||||
```scala
|
||||
import org.apache.spark.sql.expressions.Aggregator
|
||||
import org.apache.spark.sql.{Encoder, Encoders, SparkSession, functions}
|
||||
|
||||
// 1.定义员工类,对于可能存在null值的字段需要使用Option进行包装
|
||||
case class Emp(ename: String, comm: scala.Option[Double], deptno: Long, empno: Long,
|
||||
hiredate: String, job: String, mgr: scala.Option[Long], sal: Double)
|
||||
|
||||
// 2.定义聚合操作的中间输出类型
|
||||
case class SumAndCount(var sum: Double, var count: Long)
|
||||
|
||||
/* 3.自定义聚合函数
|
||||
* @IN 聚合操作的输入类型
|
||||
* @BUF reduction操作输出值的类型
|
||||
* @OUT 聚合操作的输出类型
|
||||
*/
|
||||
object MyAverage extends Aggregator[Emp, SumAndCount, Double] {
|
||||
|
||||
// 4.用于聚合操作的的初始零值
|
||||
override def zero: SumAndCount = SumAndCount(0, 0)
|
||||
|
||||
// 5.同一分区中的reduce操作
|
||||
override def reduce(avg: SumAndCount, emp: Emp): SumAndCount = {
|
||||
avg.sum += emp.sal
|
||||
avg.count += 1
|
||||
avg
|
||||
}
|
||||
|
||||
// 6.不同分区中的merge操作
|
||||
override def merge(avg1: SumAndCount, avg2: SumAndCount): SumAndCount = {
|
||||
avg1.sum += avg2.sum
|
||||
avg1.count += avg2.count
|
||||
avg1
|
||||
}
|
||||
|
||||
// 7.定义最终的输出类型
|
||||
override def finish(reduction: SumAndCount): Double = reduction.sum / reduction.count
|
||||
|
||||
// 8.中间类型的编码转换
|
||||
override def bufferEncoder: Encoder[SumAndCount] = Encoders.product
|
||||
|
||||
// 9.输出类型的编码转换
|
||||
override def outputEncoder: Encoder[Double] = Encoders.scalaDouble
|
||||
}
|
||||
|
||||
object SparkSqlApp {
|
||||
|
||||
// 测试方法
|
||||
def main(args: Array[String]): Unit = {
|
||||
|
||||
val spark = SparkSession.builder().appName("Spark-SQL").master("local[2]").getOrCreate()
|
||||
import spark.implicits._
|
||||
val ds = spark.read.json("file/emp.json").as[Emp]
|
||||
|
||||
// 10.使用内置avg()函数和自定义函数分别进行计算，验证自定义函数是否正确
|
||||
val myAvg = ds.select(MyAverage.toColumn.name("average_sal")).first()
|
||||
val avg = ds.select(functions.avg(ds.col("sal"))).first().get(0)
|
||||
|
||||
println("自定义average函数 : " + myAvg)
|
||||
println("内置的average函数 : " + avg)
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
自定义聚合函数需要实现的方法比较多，这里以绘图的方式来演示其执行流程，以及每个方法的作用：
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
关于`zero`,`reduce`,`merge`,`finish`方法的作用在上图都有说明，这里解释一下中间类型和输出类型的编码转换，这个写法比较固定，基本上就是两种情况：
|
||||
|
||||
+ 自定义类型case class或者元组就使用`Encoders.product`方法；
|
||||
+ 基本类型就使用其对应名称的方法，如`scalaByte `，`scalaFloat`，`scalaShort`等。
|
||||
|
||||
```scala
|
||||
override def bufferEncoder: Encoder[SumAndCount] = Encoders.product
|
||||
override def outputEncoder: Encoder[Double] = Encoders.scalaDouble
|
||||
```
|
||||
|
||||
|
||||
|
||||
### 6.2 无类型的自定义聚合函数
|
||||
|
||||
理解了有类型的自定义聚合函数后，无类型的定义方式也基本相同，代码如下：
|
||||
|
||||
```scala
|
||||
import org.apache.spark.sql.expressions.{MutableAggregationBuffer, UserDefinedAggregateFunction}
|
||||
import org.apache.spark.sql.types._
|
||||
import org.apache.spark.sql.{Row, SparkSession}
|
||||
|
||||
object MyAverage extends UserDefinedAggregateFunction {
|
||||
// 1.聚合操作输入参数的类型,字段名称可以自定义
|
||||
def inputSchema: StructType = StructType(StructField("MyInputColumn", LongType) :: Nil)
|
||||
|
||||
// 2.聚合操作中间值的类型,字段名称可以自定义
|
||||
def bufferSchema: StructType = {
|
||||
StructType(StructField("sum", LongType) :: StructField("MyCount", LongType) :: Nil)
|
||||
}
|
||||
|
||||
// 3.聚合操作输出参数的类型
|
||||
def dataType: DataType = DoubleType
|
||||
|
||||
// 4.此函数是否始终在相同输入上返回相同的输出,通常为true
|
||||
def deterministic: Boolean = true
|
||||
|
||||
// 5.定义零值
|
||||
def initialize(buffer: MutableAggregationBuffer): Unit = {
|
||||
buffer(0) = 0L
|
||||
buffer(1) = 0L
|
||||
}
|
||||
|
||||
// 6.同一分区中的reduce操作
|
||||
def update(buffer: MutableAggregationBuffer, input: Row): Unit = {
|
||||
if (!input.isNullAt(0)) {
|
||||
buffer(0) = buffer.getLong(0) + input.getLong(0)
|
||||
buffer(1) = buffer.getLong(1) + 1
|
||||
}
|
||||
}
|
||||
|
||||
// 7.不同分区中的merge操作
|
||||
def merge(buffer1: MutableAggregationBuffer, buffer2: Row): Unit = {
|
||||
buffer1(0) = buffer1.getLong(0) + buffer2.getLong(0)
|
||||
buffer1(1) = buffer1.getLong(1) + buffer2.getLong(1)
|
||||
}
|
||||
|
||||
// 8.计算最终的输出值
|
||||
def evaluate(buffer: Row): Double = buffer.getLong(0).toDouble / buffer.getLong(1)
|
||||
}
|
||||
|
||||
object SparkSqlApp {
|
||||
|
||||
// 测试方法
|
||||
def main(args: Array[String]): Unit = {
|
||||
|
||||
val spark = SparkSession.builder().appName("Spark-SQL").master("local[2]").getOrCreate()
|
||||
// 9.注册自定义的聚合函数
|
||||
spark.udf.register("myAverage", MyAverage)
|
||||
|
||||
val df = spark.read.json("file/emp.json")
|
||||
df.createOrReplaceTempView("emp")
|
||||
|
||||
// 10.使用自定义函数和内置函数分别进行计算
|
||||
val myAvg = spark.sql("SELECT myAverage(sal) as avg_sal FROM emp").first()
|
||||
val avg = spark.sql("SELECT avg(sal) as avg_sal FROM emp").first()
|
||||
|
||||
println("自定义average函数 : " + myAvg)
|
||||
println("内置的average函数 : " + avg)
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 参考资料
|
||||
|
||||
[Spark SQL, DataFrames and Datasets Guide > Getting Started](https://spark.apache.org/docs/latest/sql-getting-started.html)
|
||||
321
notes/SparkSQL常用聚合函数.md
Normal file
321
notes/SparkSQL常用聚合函数.md
Normal file
@ -0,0 +1,321 @@
|
||||
# 聚合函数Aggregations
|
||||
|
||||
## 一、简单聚合
|
||||
|
||||
### 1.1 数据准备
|
||||
|
||||
```scala
|
||||
// 需要导入spark sql内置的函数包
|
||||
import org.apache.spark.sql.functions._
|
||||
|
||||
val spark = SparkSession.builder().appName("aggregations").master("local[2]").getOrCreate()
|
||||
val empDF = spark.read.json("/usr/file/json/emp.json")
|
||||
// 注册为临时视图，用于后面演示SQL查询
|
||||
empDF.createOrReplaceTempView("emp")
|
||||
empDF.show()
|
||||
```
|
||||
|
||||
> 注：emp.json可以在本仓库的resources目录进行下载。
|
||||
|
||||
### 1.2 count
|
||||
|
||||
```scala
|
||||
// 计算员工人数
|
||||
empDF.select(count("ename")).show()
|
||||
```
|
||||
|
||||
### 1.3 countDistinct
|
||||
|
||||
```scala
|
||||
// 计算姓名不重复的员工人数
|
||||
empDF.select(countDistinct("deptno")).show()
|
||||
```
|
||||
|
||||
### 1.4 approx_count_distinct
|
||||
|
||||
通常在使用大型数据集时，你可能关注的只是近似值而不是准确值，这时可以使用approx_count_distinct函数，并可以使用第二个参数指定最大允许误差。
|
||||
|
||||
```scala
|
||||
empDF.select(approx_count_distinct ("ename",0.1)).show()
|
||||
```
|
||||
|
||||
### 1.5 first & last
|
||||
|
||||
获取DataFrame中指定列的第一个值或者最后一个值。
|
||||
|
||||
```scala
|
||||
empDF.select(first("ename"),last("job")).show()
|
||||
```
|
||||
|
||||
### 1.6 min & max
|
||||
|
||||
获取DataFrame中指定列的最小值或者最大值。
|
||||
|
||||
```scala
|
||||
empDF.select(min("sal"),max("sal")).show()
|
||||
```
|
||||
|
||||
### 1.7 sum & sumDistinct
|
||||
|
||||
求和以及求指定列所有不相同的值的和。
|
||||
|
||||
```scala
|
||||
empDF.select(sum("sal")).show()
|
||||
empDF.select(sumDistinct("sal")).show()
|
||||
```
|
||||
|
||||
### 1.8 avg
|
||||
|
||||
内置的求平均数的函数。
|
||||
|
||||
```scala
|
||||
empDF.select(avg("sal")).show()
|
||||
```
|
||||
|
||||
### 1.9 数学函数
|
||||
|
||||
Spark SQL中还支持多种数学聚合函数，用于通常的数学计算，以下是一些常用的例子：
|
||||
|
||||
```scala
|
||||
// 1.计算总体方差、均方差、总体标准差、样本标准差
|
||||
empDF.select(var_pop("sal"), var_samp("sal"), stddev_pop("sal"), stddev_samp("sal")).show()
|
||||
|
||||
// 2.计算偏度和峰度
|
||||
empDF.select(skewness("sal"), kurtosis("sal")).show()
|
||||
|
||||
// 3. 计算两列的皮尔逊相关系数、样本协方差、总体协方差。(这里只是演示，员工编号和薪资两列实际上并没有什么关联关系)
|
||||
empDF.select(corr("empno", "sal"), covar_samp("empno", "sal"),covar_pop("empno", "sal")).show()
|
||||
```
|
||||
|
||||
### 1.10 聚合数据到集合
|
||||
|
||||
```scala
|
||||
scala> empDF.agg(collect_set("job"), collect_list("ename")).show()
|
||||
|
||||
输出：
|
||||
+--------------------+--------------------+
|
||||
| collect_set(job)| collect_list(ename)|
|
||||
+--------------------+--------------------+
|
||||
|[MANAGER, SALESMA...|[SMITH, ALLEN, WA...|
|
||||
+--------------------+--------------------+
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 二、分组聚合
|
||||
|
||||
### 2.1 简单分组
|
||||
|
||||
```scala
|
||||
empDF.groupBy("deptno", "job").count().show()
|
||||
//等价SQL
|
||||
spark.sql("SELECT deptno, job, count(*) FROM emp GROUP BY deptno, job").show()
|
||||
|
||||
输出：
|
||||
+------+---------+-----+
|
||||
|deptno| job|count|
|
||||
+------+---------+-----+
|
||||
| 10|PRESIDENT| 1|
|
||||
| 30| CLERK| 1|
|
||||
| 10| MANAGER| 1|
|
||||
| 30| MANAGER| 1|
|
||||
| 20| CLERK| 2|
|
||||
| 30| SALESMAN| 4|
|
||||
| 20| ANALYST| 2|
|
||||
| 10| CLERK| 1|
|
||||
| 20| MANAGER| 1|
|
||||
+------+---------+-----+
|
||||
```
|
||||
|
||||
### 2.2 分组聚合
|
||||
|
||||
```scala
|
||||
empDF.groupBy("deptno").agg(count("ename").alias("人数"), sum("sal").alias("总工资")).show()
|
||||
// 等价语法
|
||||
empDF.groupBy("deptno").agg("ename"->"count","sal"->"sum").show()
|
||||
// 等价SQL
|
||||
spark.sql("SELECT deptno, count(ename) ,sum(sal) FROM emp GROUP BY deptno").show()
|
||||
|
||||
输出：
|
||||
+------+----+------+
|
||||
|deptno|人数|总工资|
|
||||
+------+----+------+
|
||||
| 10| 3|8750.0|
|
||||
| 30| 6|9400.0|
|
||||
| 20| 5|9375.0|
|
||||
+------+----+------+
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 三、自定义聚合函数
|
||||
|
||||
Scala提供了两种自定义聚合函数的方法，分别如下：
|
||||
|
||||
- 有类型的自定义聚合函数，主要适用于DataSets；
|
||||
- 无类型的自定义聚合函数，主要适用于DataFrames。
|
||||
|
||||
以下分别使用两种方式来自定义一个求平均值的聚合函数，这里以计算员工平均工资为例。两种自定义方式分别如下：
|
||||
|
||||
### 3.1 有类型的自定义函数
|
||||
|
||||
```scala
|
||||
import org.apache.spark.sql.expressions.Aggregator
|
||||
import org.apache.spark.sql.{Encoder, Encoders, SparkSession, functions}
|
||||
|
||||
// 1.定义员工类,对于可能存在null值的字段需要使用Option进行包装
|
||||
case class Emp(ename: String, comm: scala.Option[Double], deptno: Long, empno: Long,
|
||||
hiredate: String, job: String, mgr: scala.Option[Long], sal: Double)
|
||||
|
||||
// 2.定义聚合操作的中间输出类型
|
||||
case class SumAndCount(var sum: Double, var count: Long)
|
||||
|
||||
/* 3.自定义聚合函数
|
||||
* @IN 聚合操作的输入类型
|
||||
* @BUF reduction操作输出值的类型
|
||||
* @OUT 聚合操作的输出类型
|
||||
*/
|
||||
object MyAverage extends Aggregator[Emp, SumAndCount, Double] {
|
||||
|
||||
// 4.用于聚合操作的的初始零值
|
||||
override def zero: SumAndCount = SumAndCount(0, 0)
|
||||
|
||||
// 5.同一分区中的reduce操作
|
||||
override def reduce(avg: SumAndCount, emp: Emp): SumAndCount = {
|
||||
avg.sum += emp.sal
|
||||
avg.count += 1
|
||||
avg
|
||||
}
|
||||
|
||||
// 6.不同分区中的merge操作
|
||||
override def merge(avg1: SumAndCount, avg2: SumAndCount): SumAndCount = {
|
||||
avg1.sum += avg2.sum
|
||||
avg1.count += avg2.count
|
||||
avg1
|
||||
}
|
||||
|
||||
// 7.定义最终的输出类型
|
||||
override def finish(reduction: SumAndCount): Double = reduction.sum / reduction.count
|
||||
|
||||
// 8.中间类型的编码转换
|
||||
override def bufferEncoder: Encoder[SumAndCount] = Encoders.product
|
||||
|
||||
// 9.输出类型的编码转换
|
||||
override def outputEncoder: Encoder[Double] = Encoders.scalaDouble
|
||||
}
|
||||
|
||||
object SparkSqlApp {
|
||||
|
||||
// 测试方法
|
||||
def main(args: Array[String]): Unit = {
|
||||
|
||||
val spark = SparkSession.builder().appName("Spark-SQL").master("local[2]").getOrCreate()
|
||||
import spark.implicits._
|
||||
val ds = spark.read.json("file/emp.json").as[Emp]
|
||||
|
||||
// 10.使用内置avg()函数和自定义函数分别进行计算，验证自定义函数是否正确
|
||||
val myAvg = ds.select(MyAverage.toColumn.name("average_sal")).first()
|
||||
val avg = ds.select(functions.avg(ds.col("sal"))).first().get(0)
|
||||
|
||||
println("自定义average函数 : " + myAvg)
|
||||
println("内置的average函数 : " + avg)
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
自定义聚合函数需要实现的方法比较多，这里以绘图的方式来演示其执行流程，以及每个方法的作用：
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
关于`zero`,`reduce`,`merge`,`finish`方法的作用在上图都有说明，这里解释一下中间类型和输出类型的编码转换，这个写法比较固定，基本上就是两种情况：
|
||||
|
||||
- 自定义类型case class或者元组就使用`Encoders.product`方法；
|
||||
- 基本类型就使用其对应名称的方法，如`scalaByte `，`scalaFloat`，`scalaShort`等。
|
||||
|
||||
```scala
|
||||
override def bufferEncoder: Encoder[SumAndCount] = Encoders.product
|
||||
override def outputEncoder: Encoder[Double] = Encoders.scalaDouble
|
||||
```
|
||||
|
||||
|
||||
|
||||
### 3.2 无类型的自定义聚合函数
|
||||
|
||||
理解了有类型的自定义聚合函数后，无类型的定义方式也基本相同，代码如下：
|
||||
|
||||
```scala
|
||||
import org.apache.spark.sql.expressions.{MutableAggregationBuffer, UserDefinedAggregateFunction}
|
||||
import org.apache.spark.sql.types._
|
||||
import org.apache.spark.sql.{Row, SparkSession}
|
||||
|
||||
object MyAverage extends UserDefinedAggregateFunction {
|
||||
// 1.聚合操作输入参数的类型,字段名称可以自定义
|
||||
def inputSchema: StructType = StructType(StructField("MyInputColumn", LongType) :: Nil)
|
||||
|
||||
// 2.聚合操作中间值的类型,字段名称可以自定义
|
||||
def bufferSchema: StructType = {
|
||||
StructType(StructField("sum", LongType) :: StructField("MyCount", LongType) :: Nil)
|
||||
}
|
||||
|
||||
// 3.聚合操作输出参数的类型
|
||||
def dataType: DataType = DoubleType
|
||||
|
||||
// 4.此函数是否始终在相同输入上返回相同的输出,通常为true
|
||||
def deterministic: Boolean = true
|
||||
|
||||
// 5.定义零值
|
||||
def initialize(buffer: MutableAggregationBuffer): Unit = {
|
||||
buffer(0) = 0L
|
||||
buffer(1) = 0L
|
||||
}
|
||||
|
||||
// 6.同一分区中的reduce操作
|
||||
def update(buffer: MutableAggregationBuffer, input: Row): Unit = {
|
||||
if (!input.isNullAt(0)) {
|
||||
buffer(0) = buffer.getLong(0) + input.getLong(0)
|
||||
buffer(1) = buffer.getLong(1) + 1
|
||||
}
|
||||
}
|
||||
|
||||
// 7.不同分区中的merge操作
|
||||
def merge(buffer1: MutableAggregationBuffer, buffer2: Row): Unit = {
|
||||
buffer1(0) = buffer1.getLong(0) + buffer2.getLong(0)
|
||||
buffer1(1) = buffer1.getLong(1) + buffer2.getLong(1)
|
||||
}
|
||||
|
||||
// 8.计算最终的输出值
|
||||
def evaluate(buffer: Row): Double = buffer.getLong(0).toDouble / buffer.getLong(1)
|
||||
}
|
||||
|
||||
object SparkSqlApp {
|
||||
|
||||
// 测试方法
|
||||
def main(args: Array[String]): Unit = {
|
||||
|
||||
val spark = SparkSession.builder().appName("Spark-SQL").master("local[2]").getOrCreate()
|
||||
// 9.注册自定义的聚合函数
|
||||
spark.udf.register("myAverage", MyAverage)
|
||||
|
||||
val df = spark.read.json("file/emp.json")
|
||||
df.createOrReplaceTempView("emp")
|
||||
|
||||
// 10.使用自定义函数和内置函数分别进行计算
|
||||
val myAvg = spark.sql("SELECT myAverage(sal) as avg_sal FROM emp").first()
|
||||
val avg = spark.sql("SELECT avg(sal) as avg_sal FROM emp").first()
|
||||
|
||||
println("自定义average函数 : " + myAvg)
|
||||
println("内置的average函数 : " + avg)
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 参考资料
|
||||
|
||||
1. Matei Zaharia, Bill Chambers . Spark: The Definitive Guide[M] . 2018-02
|
||||
|
||||
|
||||
|
||||
14
notes/SparkSQL数据源支持.md
Normal file
14
notes/SparkSQL数据源支持.md
Normal file
@ -0,0 +1,14 @@
|
||||
1.1 Json
|
||||
|
||||
```scala
|
||||
val empDF = spark.read.json("/usr/file/json/emp.json")
|
||||
empDF.show()
|
||||
```
|
||||
|
||||
1.2
|
||||
|
||||
```scala
|
||||
val parquetFileDF = spark.read.parquet("/usr/file/parquet/emp.parquet")
|
||||
parquetFileDF.show()
|
||||
```
|
||||
|
||||
182
notes/Spark_Structured_API的基本使用.md
Normal file
182
notes/Spark_Structured_API的基本使用.md
Normal file
@ -0,0 +1,182 @@
|
||||
# Structured API基本使用
|
||||
|
||||
## 一、创建DataFrames
|
||||
|
||||
Spark中所有功能的入口点是`SparkSession`，可以使用`SparkSession.builder()`创建。创建后应用程序就可以从现有RDD，Hive表或Spark数据源创建DataFrame。如下所示：
|
||||
|
||||
```scala
|
||||
val spark = SparkSession.builder().appName("Spark-SQL").master("local[2]").getOrCreate()
|
||||
val df = spark.read.json("/usr/file/emp.json")
|
||||
df.show()
|
||||
|
||||
// 建议在进行spark SQL编程前导入下面的隐式转换，因为DataFrames和dataSets中很多操作都依赖了隐式转换
|
||||
import spark.implicits._
|
||||
```
|
||||
|
||||
这里可以启动`spark-shell`进行测试，需要注意的是`spark-shell`启动后会自动创建一个名为`spark`的`SparkSession`，在命令行中可以直接引用即可：
|
||||
|
||||
|
||||
|
||||
## 二、DataFrames基本操作
|
||||
|
||||
### 2.1 printSchema
|
||||
|
||||
```scala
|
||||
// 以树形结构打印dataframe的schema信息
|
||||
df.printSchema()
|
||||
```
|
||||
|
||||
|
||||
|
||||
### 2.2 使用DataFrame API进行基本查询
|
||||
|
||||
```scala
|
||||
// 查询员工姓名及工作
|
||||
df.select($"ename", $"job").show()
|
||||
|
||||
// 查询工资大于2000的员工信息
|
||||
df.filter($"sal" > 2000).show()
|
||||
|
||||
// 分组统计部门人数
|
||||
df.groupBy("deptno").count().show()
|
||||
```
|
||||
|
||||
### 2.3 使用SQL进行基本查询
|
||||
|
||||
```scala
|
||||
// 首先需要将DataFrame注册为临时视图
|
||||
df.createOrReplaceTempView("emp")
|
||||
|
||||
// 查询员工姓名及工作
|
||||
spark.sql("SELECT ename,job FROM emp").show()
|
||||
|
||||
// 查询工资大于2000的员工信息
|
||||
spark.sql("SELECT * FROM emp where sal > 2000").show()
|
||||
|
||||
// 分组统计部门人数
|
||||
spark.sql("SELECT deptno,count(ename) FROM emp group by deptno").show()
|
||||
```
|
||||
|
||||
### 2.4 全局临时视图
|
||||
|
||||
上面使用`createOrReplaceTempView`创建的是会话临时视图，它的生命周期仅限于会话范围，会随会话的结束而结束。
|
||||
|
||||
你也可以使用`createGlobalTempView`创建全局临时视图，全局临时视图可以在所有会话之间共享，并直到整个Spark应用程序终止才会消失。全局临时视图被定义在内置的`global_temp`数据库下，需要使用限定名称进行引用，如`SELECT * FROM global_temp.view1`。
|
||||
|
||||
```scala
|
||||
// 注册为全局临时视图
|
||||
df.createGlobalTempView("gemp")
|
||||
|
||||
// 查询员工姓名及工作，使用限定名称进行引用
|
||||
spark.sql("SELECT ename,job FROM global_temp.gemp").show()
|
||||
|
||||
// 查询工资大于2000的员工信息，使用限定名称进行引用
|
||||
spark.sql("SELECT * FROM global_temp.gemp where sal > 2000").show()
|
||||
|
||||
// 分组统计部门人数，使用限定名称进行引用
|
||||
spark.sql("SELECT deptno,count(ename) FROM global_temp.gemp group by deptno").show()
|
||||
```
|
||||
|
||||
## 三、创建Datasets
|
||||
|
||||
### 3.1 由外部数据集创建
|
||||
|
||||
```scala
|
||||
// 1.需要导入隐式转换
|
||||
import spark.implicits._
|
||||
|
||||
// 2.创建case class,等价于Java Bean
|
||||
case class Emp(ename: String, comm: Double, deptno: Long, empno: Long,
|
||||
hiredate: String, job: String, mgr: Long, sal: Double)
|
||||
|
||||
// 3.由外部数据集创建Datasets
|
||||
val ds = spark.read.json("/usr/file/emp.json").as[Emp]
|
||||
ds.show()
|
||||
```
|
||||
|
||||
### 3.2 由内部数据集创建
|
||||
|
||||
```scala
|
||||
// 1.需要导入隐式转换
|
||||
import spark.implicits._
|
||||
|
||||
// 2.创建case class,等价于Java Bean
|
||||
case class Emp(ename: String, comm: Double, deptno: Long, empno: Long,
|
||||
hiredate: String, job: String, mgr: Long, sal: Double)
|
||||
|
||||
// 3.由内部数据集创建Datasets
|
||||
val caseClassDS = Seq(Emp("ALLEN", 300.0, 30, 7499, "1981-02-20 00:00:00", "SALESMAN", 7698, 1600.0),
|
||||
Emp("JONES", 300.0, 30, 7499, "1981-02-20 00:00:00", "SALESMAN", 7698, 1600.0))
|
||||
.toDS()
|
||||
caseClassDS.show()
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 四、DataFrames与Datasets互相转换
|
||||
|
||||
Spark提供了非常简单的转换方法用于DataFrames与Datasets互相转换，示例如下：
|
||||
|
||||
```shell
|
||||
# DataFrames转Datasets
|
||||
scala> df.as[Emp]
|
||||
res1: org.apache.spark.sql.Dataset[Emp] = [COMM: double, DEPTNO: bigint ... 6 more fields]
|
||||
|
||||
# Datasets转DataFrames
|
||||
scala> ds.toDF()
|
||||
res2: org.apache.spark.sql.DataFrame = [COMM: double, DEPTNO: bigint ... 6 more fields]
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 五、RDDs转换为DataFrames\Datasets
|
||||
|
||||
Spark支持两种方式把RDD转换为DataFrames，分别是使用反射推断和指定schema转换。
|
||||
|
||||
### 5.1 使用反射推断
|
||||
|
||||
```scala
|
||||
// 1.导入隐式转换
|
||||
import spark.implicits._
|
||||
|
||||
// 2.创建部门类
|
||||
case class Dept(deptno: Long, dname: String, loc: String)
|
||||
|
||||
// 3.创建RDD并转换为dataSet
|
||||
val rddToDS = spark.sparkContext
|
||||
.textFile("/usr/file/dept.txt")
|
||||
.map(_.split("\t"))
|
||||
.map(line => Dept(line(0).trim.toLong, line(1), line(2)))
|
||||
.toDS() // 如果调用toDF()则转换为dataFrame
|
||||
```
|
||||
|
||||
### 5.2 以编程方式指定Schema
|
||||
|
||||
```scala
|
||||
import org.apache.spark.sql.Row
|
||||
import org.apache.spark.sql.types._
|
||||
|
||||
|
||||
// 1.定义每个列的列类型
|
||||
val fields = Array(StructField("deptno", LongType, nullable = true),
|
||||
StructField("dname", StringType, nullable = true),
|
||||
StructField("loc", StringType, nullable = true))
|
||||
|
||||
// 2.创建schema
|
||||
val schema = StructType(fields)
|
||||
|
||||
// 3.创建RDD
|
||||
val deptRDD = spark.sparkContext.textFile("/usr/file/dept.txt")
|
||||
val rowRDD = deptRDD.map(_.split("\t")).map(line => Row(line(0).toLong, line(1), line(2)))
|
||||
|
||||
|
||||
// 4.将RDD转换为dataFrame
|
||||
val deptDF = spark.createDataFrame(rowRDD, schema)
|
||||
deptDF.show()
|
||||
```
|
||||
|
||||
|
||||
|
||||
## 参考资料
|
||||
|
||||
[Spark SQL, DataFrames and Datasets Guide > Getting Started](https://spark.apache.org/docs/latest/sql-getting-started.html)
|
||||
BIN
resources/parquet/dept.parquet/._SUCCESS.crc
Normal file
BIN
resources/parquet/dept.parquet/._SUCCESS.crc
Normal file
Binary file not shown.
Binary file not shown.
0
resources/parquet/dept.parquet/_SUCCESS
Normal file
0
resources/parquet/dept.parquet/_SUCCESS
Normal file
Binary file not shown.
BIN
resources/parquet/emp.parquet/._SUCCESS.crc
Normal file
BIN
resources/parquet/emp.parquet/._SUCCESS.crc
Normal file
Binary file not shown.
Binary file not shown.
0
resources/parquet/emp.parquet/_SUCCESS
Normal file
0
resources/parquet/emp.parquet/_SUCCESS
Normal file
Binary file not shown.
4
resources/txt/dept.txt
Normal file
4
resources/txt/dept.txt
Normal file
@ -0,0 +1,4 @@
|
||||
10 ACCOUNTING NEW YORK
|
||||
20 RESEARCH DALLAS
|
||||
30 SALES CHICAGO
|
||||
40 OPERATIONS BOSTON
|
||||
14
resources/txt/emp.txt
Normal file
14
resources/txt/emp.txt
Normal file
@ -0,0 +1,14 @@
|
||||
7369 SMITH CLERK 7902 1980-12-17 00:00:00 800.00 20
|
||||
7499 ALLEN SALESMAN 7698 1981-02-20 00:00:00 1600.00 300.00 30
|
||||
7521 WARD SALESMAN 7698 1981-02-22 00:00:00 1250.00 500.00 30
|
||||
7566 JONES MANAGER 7839 1981-04-02 00:00:00 2975.00 20
|
||||
7654 MARTIN SALESMAN 7698 1981-09-28 00:00:00 1250.00 1400.00 30
|
||||
7698 BLAKE MANAGER 7839 1981-05-01 00:00:00 2850.00 30
|
||||
7782 CLARK MANAGER 7839 1981-06-09 00:00:00 2450.00 10
|
||||
7788 SCOTT ANALYST 7566 1987-04-19 00:00:00 1500.00 20
|
||||
7839 KING PRESIDENT 1981-11-17 00:00:00 5000.00 10
|
||||
7844 TURNER SALESMAN 7698 1981-09-08 00:00:00 1500.00 0.00 30
|
||||
7876 ADAMS CLERK 7788 1987-05-23 00:00:00 1100.00 20
|
||||
7900 JAMES CLERK 7698 1981-12-03 00:00:00 950.00 30
|
||||
7902 FORD ANALYST 7566 1981-12-03 00:00:00 3000.00 20
|
||||
7934 MILLER CLERK 7782 1982-01-23 00:00:00 1300.00 10
|
||||
Loading…
x
Reference in New Issue
Block a user