基本概念
What's DataFrame
A DataFrame is equivalent to a relational table in Spark SQL [1]。
DataFrame的前身是SchemaRDD，从Spark 1.3.0开始SchemaRDD更名为DataFrame [2]。其实从使用上来看，跟RDD的区别主要是有了Schema，这样就能根据不同行和列得到对应的值。
Why DataFrame, Motivition
比RDD有更多的操作，而且执行计划上也比RDD有更多的优化。能够方便处理大规模结构化数据。
How to use DataFrame
创建DataFrame
- 创建一个空的DataFrame
这里schema是一个StructType类型的
sqlContext.createDataFrame(sc.emptyRDD[Row], schema)
- 从一个List创建
def listToDataFrame(list: ListBuffer[List[Any]], schema:StructType): DataFrame = {
val rows = list.map{x => Row(x:_*)}
val rdd = sqlContext.sparkContext.parallelize(rows)
sqlContext.createDataFrame(rdd, schema)
}
- 直接通过RDD生成
val departments = sc.parallelize(Array(
(31, "Sales"),
(33, "Engineering"),
(34, "Clerical"),
(35, "Marketing")
)).toDF("DepartmentID", "DepartmentName")
val employees = sc.parallelize(Array[(String, Option[Int])](
("Rafferty", Some(31)), ("Jones", Some(33)), ("Heisenberg", Some(33)), ("Robinson", Some(34)), ("Smith", Some(34)),
("Williams", null)
)).toDF("LastName", "DepartmentID")
- 读取json文件创建[5]
json文件
{"name":"Michael"}
{"name":"Andy", "age":30}
{"name":"Justin", "age":19}
创建DataFrame
val df = sqlContext.jsonFile("/path/to/your/jsonfile")
df: org.apache.spark.sql.DataFrame = [age: bigint, name: string]
- 从parquet文件读出创建
val df:DataFrame = sqlContext.read.parquet("/Users/robin/workspace/cooked_data/bt")
- 从MySQL读取表chuang创建[5]
val jdbcDF = sqlContext.load("jdbc", Map("url" -> "jdbc:mysql://localhost:3306/db?user=aaa&password=111", "dbtable" -> "your_table"))
- 从Hive创建[5]
Spark提供了一个HiveContext的上下文，其实是SQLContext的一个子类，但从作用上来说，sqlContext也支持Hive数据源。只要在部署Spark的时候加入Hive选项，并把已有的hive-site.xml文件挪到$SPARK_HOME/conf路径下，就可以直接用Spark查询包含已有元数据的Hive表了
sqlContext.sql("select count(*) from hive_people")
- 从CSV文件创建
有个spark-csv的library
可以从maven引入，也可以k从spark-shell $SPARK_HOME/bin/spark-shell --packages com.databricks:spark-csv_2.11:1.5.0
val df = sqlContext.read.format("com.databricks.spark.csv").
option("header", "true").
option("inferSchema","true").
load("/Users/username/tmp/person.csv")
DataFrame基本操作
官方例子
// To create DataFrame using SQLContext
val 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")))
Filter
- 把id为null的行都filter掉
df.withColumn("id", when(expr("id is null"), 0).otherwise(1)).show
Join连接
- inner join [4]
val employees = sc.parallelize(Array[(String, Option[Int])](
("Rafferty", Some(31)), ("Jones", Some(33)), ("Heisenberg", Some(33)), ("Robinson", Some(34)), ("Smith", Some(34)), ("Williams", null)
)).toDF("LastName", "DepartmentID")
val departments = sc.parallelize(Array(
(31, "Sales"), (33, "Engineering"), (34, "Clerical"),
(35, "Marketing")
)).toDF("DepartmentID", "DepartmentName")
departments.show()
+------------+--------------+
|DepartmentID|DepartmentName|
+------------+--------------+
| 31| Sales|
| 33| Engineering|
| 34| Clerical|
| 35| Marketing|
+------------+--------------+
employees.join(departments, "DepartmentID").show()
+------------+----------+--------------+
|DepartmentID| LastName|DepartmentName|
+------------+----------+--------------+
| 31| Rafferty| Sales|
| 33| Jones| Engineering|
| 33|Heisenberg| Engineering|
| 34| Robinson| Clerical|
| 34| Smith| Clerical|
| null| Williams| null|
+------------+----------+--------------+
- left outer join [4]
employees.join(departments, Seq("DepartmentID"), "left_outer").show()
+------------+----------+--------------+
|DepartmentID| LastName|DepartmentName|
+------------+----------+--------------+
| 31| Rafferty| Sales|
| 33| Jones| Engineering|
| 33|Heisenberg| Engineering|
| 34| Robinson| Clerical|
| 34| Smith| Clerical|
| null| Williams| null|
+------------+----------+--------------+
val d1 = df.groupBy("startDate","endDate").agg(max("price") as "price").show
- Join expression 用表达式连接 [3]
val products = sc.parallelize(Array(
("steak", "1990-01-01", "2000-01-01", 150),
("steak", "2000-01-02", "2020-01-01", 180),
("fish", "1990-01-01", "2020-01-01", 100)
)).toDF("name", "startDate", "endDate", "price")
products.show()
+-----+----------+----------+-----+
| name| startDate| endDate|price|
+-----+----------+----------+-----+
|steak|1990-01-01|2000-01-01| 150|
|steak|2000-01-02|2020-01-01| 180|
| fish|1990-01-01|2020-01-01| 100|
+-----+----------+----------+-----+
val orders = sc.parallelize(Array(
("1995-01-01", "steak"),
("2000-01-01", "fish"),
("2005-01-01", "steak")
)).toDF("date", "product")
orders.show()
+----------+-------+
| date|product|
+----------+-------+
|1995-01-01| steak|
|2000-01-01| fish|
|2005-01-01| steak|
+----------+-------+
orders.join(products, $"product" === $"name" && $"date" >= $"startDate" && $"date" <= $"endDate") .show()
+----------+-------+-----+----------+----------+-----+
| date|product| name| startDate| endDate|price|
+----------+-------+-----+----------+----------+-----+
|2000-01-01| fish| fish|1990-01-01|2020-01-01| 100|
|1995-01-01| steak|steak|1990-01-01|2000-01-01| 150|
|2005-01-01| steak|steak|2000-01-02|2020-01-01| 180|
+----------+-------+-----+----------+----------+-----+
- Join types:
inner, outer, left_outer, right_outer, leftsemi - Join with dataframe alias
val joinedDF = testDF.as('a).join(genmodDF.as('b), $"a.PassengerId" === $"b.PassengerId")
joinedDF.select($"a.PassengerId", $"b.PassengerId").take(10)
val joinedDF = testDF.join(genmodDF, testDF("PassengerId") === genmodDF("PassengerId"), "inner")
Reference
- [1] 官方API文档 https://spark.apache.org/docs/1.6.3/api/java/index.html?org/apache/spark/sql/DataFrame.htm
- [2] spark结构化数据处理：Spark SQL、DataFrame和Dataset http://weibo.com/ttarticle/p/show?id=2309403994211933363947&sudaref=www.google.com&retcode=6102
- [3] Joining Data Frames in Spark SQL http://bailiwick.io/2015/07/13/joining-data-frames-in-spark-sql/
- [4] Beyond traditional join with Apache Spark http://kirillpavlov.com/blog/2016/04/23/beyond-traditional-join-with-apache-spark/
- [5] Spark DataFrame小试牛刀 http://guoze.me/2015/03/22/spark-dataframe/