Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

Unable to create dataframe from RDD of Row using case class

Using Spark 2.x, it seems I cannot create a Dataframe using a RDD of Row composed of case classes.

It did work fine on Spark 1.6.x but fails on 2.x the following runtime exception:

java.lang.RuntimeException: Timestamp is not a valid external type for schema of struct<seconds:bigint,nanos:int>

preceded by a bunch of generated code from Catalyst.

Here is the snippet (simplified version of what I am doing):

package main

import org.apache.spark.sql.{SparkSession, Row}
import org.apache.spark.sql.types.{IntegerType, LongType, StructField, StructType}

object Test {

  case class Timestamp(seconds: Long, nanos: Int)

  val TIMESTAMP_TYPE = StructType(List(
    StructField("seconds", LongType, false),
    StructField("nanos", IntegerType, false)
  ))

  val SCHEMA = StructType(List(
    StructField("created_at", TIMESTAMP_TYPE, true)
  ))

  def main(args: Array[String]) {

    val spark = SparkSession.builder().getOrCreate()

    val rowRDD = spark.sparkContext.parallelize(Seq((0L, 0))).map {
      case (seconds: Long, nanos: Int) => {
        Row(Timestamp(seconds, nanos))
      }
    }

    spark.createDataFrame(rowRDD, SCHEMA).show(1)
  }
}

I am not sure if this is a Spark bug or something I missed in the documentation (I know Spark 2.x introduced runtime Row encoding validation, maybe this is related)

Help much appreciated

like image 232
cerisier Avatar asked Sep 01 '16 20:09

cerisier


1 Answers

I am not sure if it is a bug or not but mixing dynamically typed Row, case classes and explicit schema doesn't make much sense. Either use Rows and schema:

import collection.mutable._
import collection.JavaConverters._

spark.createDataFrame(ArrayBuffer(Row(Row(0L, 0))).asJava, SCHEMA)

or case classes:

import spark.implicits._

Seq(Tuple1(Timestamp(0L, 0))).toDF("created_at")

Otherwise you're just doing the same job twice.

Note:

If you want express that fields can be nullable you use Options. For example

case class Record(created_at: Option[Timestamp])
case class Timestamp(seconds: Long, nanos: Option[Int])

Seq(Record(Some(Timestamp(0L, Some(0))))).toDF

will generate schema where created_at and created_at.milliseconds can be NULL, but created_at.seconds has to be set if created_at is not NULL.

like image 138
zero323 Avatar answered Nov 15 '22 06:11

zero323