Spark: Programmatically creating dataframe schema

问题:

I have a smallish dataset that will be the result of a Spark job. I am thinking about converting this dataset to a dataframe for convenience at the end of the job, but have struggled to correctly define the schema. The problem is the last field below (topValues); it is an ArrayBuffer of tuples -- keys and counts.

  val innerSchema =
    StructType(
      Array(
        StructField("value", StringType),
        StructField("count", LongType)
      )
    )
  val outputSchema =
    StructType(
      Array(
        StructField("name", StringType, nullable=false),
        StructField("index", IntegerType, nullable=false),
        StructField("count", LongType, nullable=false),
        StructField("empties", LongType, nullable=false),
        StructField("nulls", LongType, nullable=false),
        StructField("uniqueValues", LongType, nullable=false),
        StructField("mean", DoubleType),
        StructField("min", DoubleType),
        StructField("max", DoubleType),
        StructField("topValues", innerSchema)
      )
    )

  val result = stats.columnStats.map{ c =>
    Row(c._2.name, c._1, c._2.count, c._2.empties, c._2.nulls, c._2.uniqueValues, c._2.mean, c._2.min, c._2.max, c._2.topValues.topN)
  }

  val rdd = sc.parallelize(result.toSeq)

  val outputDf = sqlContext.createDataFrame(rdd, outputSchema)

  outputDf.show()

The error I'm getting is a MatchError: scala.MatchError: ArrayBuffer((10,2), (20,3), (8,1)) (of class scala.collection.mutable.ArrayBuffer)

When I debug and inspect my objects, I'm seeing this:

rdd: ParallelCollectionRDD[2]
rdd.data: "ArrayBuffer" size = 2
rdd.data(0): [age,2,6,0,0,3,14.666666666666666,8.0,20.0,ArrayBuffer((10,2), (20,3), (8,1))]
rdd.data(1): [gender,3,6,0,0,2,0.0,0.0,0.0,ArrayBuffer((M,4), (F,2))]

It seems to me that I've accurately described the ArrayBuffer of tuples in my innerSchema, but Spark disagrees.

Any idea how I should be defining the schema?

回答1:

val rdd = sc.parallelize(Array(Row(ArrayBuffer(1,2,3,4))))
val df = sqlContext.createDataFrame(
  rdd,
  StructType(Seq(StructField("arr", ArrayType(IntegerType, false), false)
)

df.printSchema
root
 |-- arr: array (nullable = false)
 |    |-- element: integer (containsNull = false)

df.show
+------------+
|         arr|
+------------+
|[1, 2, 3, 4]|
+------------+

回答2:

As David pointed out, I needed to use an ArrayType. Spark is happy with this:

  val outputSchema =
    StructType(
      Array(
        StructField("name", StringType, nullable=false),
        StructField("index", IntegerType, nullable=false),
        StructField("count", LongType, nullable=false),
        StructField("empties", LongType, nullable=false),
        StructField("nulls", LongType, nullable=false),
        StructField("uniqueValues", LongType, nullable=false),
        StructField("mean", DoubleType),
        StructField("min", DoubleType),
        StructField("max", DoubleType),
        StructField("topValues", ArrayType(StructType(Array(
          StructField("value", StringType),
          StructField("count", LongType)
        ))))
      )
    )

回答3:

import spark.implicits._
import org.apache.spark.sql.types._
import org.apache.spark.sql.functions._


val searchPath = "/path/to/.csv"
val columns = "col1,col2,col3,col4,col5,col6,col7"
val fields = columns.split(",").map(fieldName => StructField(fieldName, StringType, 
nullable = true))
val customSchema = StructType(fields)
var dfPivot =spark.read.format("com.databricks.spark.csv").option("header","false").option("inferSchema", "false").schema(customSchema).load(searchPath)

When you load the data with custom schema will be much faster compared to loading data with default schema