How can I create a Spark DataFrame from a nested array of struct element?

Question

I have read a JSON file into Spark. This file has the following structure:

scala> tweetBlob.printSchema
root
 |-- related: struct (nullable = true)
 |    |-- next: struct (nullable = true)
 |    |    |-- href: string (nullable = true)
 |-- search: struct (nullable = true)
 |    |-- current: long (nullable = true)
 |    |-- results: long (nullable = true)
 |-- tweets: array (nullable = true)
 |    |-- element: struct (containsNull = true)
 |    |    |-- cde: struct (nullable = true)
...
...
 |    |    |-- cdeInternal: struct (nullable = true)
...
...
 |    |    |-- message: struct (nullable = true)
...
...

What I would ideally want is a DataFrame with columns "cde", "cdeInternal", "message"... as shown below

root
|-- cde: struct (nullable = true)
...
...
|-- cdeInternal: struct (nullable = true)
...
...
|-- message: struct (nullable = true)
...
...

I have managed to use "explode" to extract elements from the "tweets" array into a column called "tweets"

scala> val tweets = tweetBlob.select(explode($"tweets").as("tweets"))
tweets: org.apache.spark.sql.DataFrame = [tweets: struct<cde:struct<author:struct<gender:string,location:struct<city:string,country:string,state:string>,maritalStatus:struct<evidence:string,isMarried:string>,parenthood:struct<evidence:string,isParent:string>>,content:struct<sentiment:struct<evidence:array<struct<polarity:string,sentimentTerm:string>>,polarity:string>>>,cdeInternal:struct<compliance:struct<isActive:boolean,userProtected:boolean>,tracks:array<struct<id:string>>>,message:struct<actor:struct<displayName:string,favoritesCount:bigint,followersCount:bigint,friendsCount:bigint,id:string,image:string,languages:array<string>,link:string,links:array<struct<href:string,rel:string>>,listedCount:bigint,location:struct<displayName:string,objectType:string>,objectType:string,postedTime...
scala> tweets.printSchema
root
 |-- tweets: struct (nullable = true)
 |    |-- cde: struct (nullable = true)
...
...
 |    |-- cdeInternal: struct (nullable = true)
...
...
 |    |-- message: struct (nullable = true)
...
...

How can I select all columns inside the struct and create a DataFrame out of it? Explode does not work on a struct if my understanding is correct.

Any help is appreciated.

Duffer

Question also asked at Databricks forum, but no response. – Duffer

Answer 1

One possible way to handle this is to extract required information from the schema. Lets start with some dummy data:

import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.types._


case class Bar(x: Int, y: String)
case class Foo(bar: Bar)

val df = sc.parallelize(Seq(Foo(Bar(1, "first")), Foo(Bar(2, "second")))).toDF

df.printSchema

// root
//  |-- bar: struct (nullable = true)
//  |    |-- x: integer (nullable = false)
//  |    |-- y: string (nullable = true)

and a helper function:

def children(colname: String, df: DataFrame) = {
  val parent = df.schema.fields.filter(_.name == colname).head
  val fields = parent.dataType match {
    case x: StructType => x.fields
    case _ => Array.empty[StructField]
  }
  fields.map(x => col(s"$colname.${x.name}"))
}

Finally the results:

df.select(children("bar", df): _*).printSchema

// root
// |-- x: integer (nullable = true)
// |-- y: string (nullable = true)

Selfconsistent

Hi @Selfconsistent it might be a silly question, but what does "children("bar", df): _*" this syntax mean? – Repress

children("bar", df) is just a call which returns Seq[Column]. :_* performs varargs unpacking. – Selfconsistent

Is there a way to do this in SparkR? – Refrigerate

@Refrigerate It should be. SparkR schema is equivalent to the Scala one, and dot syntax is universal for all Spark SQL implementations. – Selfconsistent

Answer 2

You can use

df
  .select(explode(col("path_to_collection")).as("collection"))
  .select(col("collection.*"))`:

Example:

scala> val json = """{"name":"Michael", "schools":[{"sname":"stanford", "year":2010}, {"sname":"berkeley", "year":2012}]}"""

scala> val inline = sqlContext.read.json(sc.parallelize(json :: Nil)).select(explode(col("schools")).as("collection")).select(col("collection.*"))

scala> inline.printSchema
root
 |-- sname: string (nullable = true)
 |-- year: long (nullable = true)

scala> inline.show
+--------+----+
|   sname|year|
+--------+----+
|stanford|2010|
|berkeley|2012|
+--------+----+

Or, you can also use SQL function inline:

scala> val json = """{"name":"Michael", "schools":[{"sname":"stanford", "year":2010}, {"sname":"berkeley", "year":2012}]}"""

scala> sqlContext.read.json(sc.parallelize(json :: Nil)).registerTempTable("tmp")

scala> val inline = sqlContext.sql("SELECT inline(schools) FROM tmp")

scala> inline.printSchema
root
 |-- sname: string (nullable = true)
 |-- year: long (nullable = true)

scala> inline.show
+--------+----+
|   sname|year|
+--------+----+
|stanford|2010|
|berkeley|2012|
+--------+----+

Ardeth

improving your answer to w.r.t Spark 2.0. spark.read.json(spark.createDataset(json :: Nil)).createOrReplaceTempView("tmp") – Kedge

This sql query may useful to the above solution """select name,inline(schools) from tmp""" – Kedge

Answer 3

scala> import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.DataFrame

scala> import org.apache.spark.sql.types._
import org.apache.spark.sql.types._

scala> case class Bar(x: Int, y: String)
defined class Bar

scala> case class Foo(bar: Bar)
defined class Foo

scala> val df = sc.parallelize(Seq(Foo(Bar(1, "first")), Foo(Bar(2, "second")))).toDF
df: org.apache.spark.sql.DataFrame = [bar: struct<x: int, y: string>]


scala> df.printSchema
root
 |-- bar: struct (nullable = true)
 |    |-- x: integer (nullable = false)
 |    |-- y: string (nullable = true)


scala> df.select("bar.*").printSchema
root
 |-- x: integer (nullable = true)
 |-- y: string (nullable = true)


scala> 

Creamy

This method is way easier. – Auriga