Spark 2.1 cannot write Vector field on CSV

问题

I was migrating my code from Spark 2.0 to 2.1 when I stumbled into a problem related to Dataframe saving.

Here's the code

import org.apache.spark.sql.types._
import org.apache.spark.ml.linalg.VectorUDT
val df = spark.createDataFrame(Seq(Tuple1(1))).toDF("values")
val toSave = new org.apache.spark.ml.feature.VectorAssembler().setInputCols(Array("values")).transform(df)
toSave.write.csv(path)

This code succeeds when using Spark 2.0.0

Using Spark 2.1.0.cloudera1, I get the following error :

java.lang.UnsupportedOperationException: CSV data source does not support struct<type:tinyint,size:int,indices:array<int>,values:array<double>> data type.
  at org.apache.spark.sql.execution.datasources.csv.CSVFileFormat.org$apache$spark$sql$execution$datasources$csv$CSVFileFormat$$verifyType$1(CSVFileFormat.scala:233)
  at org.apache.spark.sql.execution.datasources.csv.CSVFileFormat$$anonfun$verifySchema$1.apply(CSVFileFormat.scala:237)
  at org.apache.spark.sql.execution.datasources.csv.CSVFileFormat$$anonfun$verifySchema$1.apply(CSVFileFormat.scala:237)
  at scala.collection.Iterator$class.foreach(Iterator.scala:893)
  at scala.collection.AbstractIterator.foreach(Iterator.scala:1336)
  at scala.collection.IterableLike$class.foreach(IterableLike.scala:72)
  at org.apache.spark.sql.types.StructType.foreach(StructType.scala:96)
  at org.apache.spark.sql.execution.datasources.csv.CSVFileFormat.verifySchema(CSVFileFormat.scala:237)
  at org.apache.spark.sql.execution.datasources.csv.CSVFileFormat.prepareWrite(CSVFileFormat.scala:121)
  at org.apache.spark.sql.execution.datasources.FileFormatWriter$.write(FileFormatWriter.scala:108)
  at org.apache.spark.sql.execution.datasources.InsertIntoHadoopFsRelationCommand.run(InsertIntoHadoopFsRelationCommand.scala:101)
  at org.apache.spark.sql.execution.command.ExecutedCommandExec.sideEffectResult$lzycompute(commands.scala:58)
  at org.apache.spark.sql.execution.command.ExecutedCommandExec.sideEffectResult(commands.scala:56)
  at org.apache.spark.sql.execution.command.ExecutedCommandExec.doExecute(commands.scala:74)
  at org.apache.spark.sql.execution.SparkPlan$$anonfun$execute$1.apply(SparkPlan.scala:114)
  at org.apache.spark.sql.execution.SparkPlan$$anonfun$execute$1.apply(SparkPlan.scala:114)
  at org.apache.spark.sql.execution.SparkPlan$$anonfun$executeQuery$1.apply(SparkPlan.scala:135)
  at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:151)
  at org.apache.spark.sql.execution.SparkPlan.executeQuery(SparkPlan.scala:132)
  at org.apache.spark.sql.execution.SparkPlan.execute(SparkPlan.scala:113)
  at org.apache.spark.sql.execution.QueryExecution.toRdd$lzycompute(QueryExecution.scala:87)
  at org.apache.spark.sql.execution.QueryExecution.toRdd(QueryExecution.scala:87)
  at org.apache.spark.sql.execution.datasources.DataSource.writeInFileFormat(DataSource.scala:484)
  at org.apache.spark.sql.execution.datasources.DataSource.write(DataSource.scala:520)
  at org.apache.spark.sql.DataFrameWriter.save(DataFrameWriter.scala:215)
  at org.apache.spark.sql.DataFrameWriter.save(DataFrameWriter.scala:198)
  at org.apache.spark.sql.DataFrameWriter.csv(DataFrameWriter.scala:579)
  ... 50 elided

Is this only on my side ?

Is this related to the cloudera release of Spark 2.1 ? (from their repo, it seems they didn't mess with spark.sql so maybe not)

Thanks !

回答1:

The following answer is composed from @zero323's comment.

CSV source doesn't support complex objects. Exactly as you from the exception: CSV data source does not support struct,values:array‌​> data type. is an expected behavior. It doesn't work with Spark 2.x although it used to work with spark-csv in 1.x where vectors have been converted to strings.

This behavior was correct in the following jira SPARK-16216.

回答2:

As a workaround you can use the VectorDisassembler class from this fork, or take the solution described here.

I have used VectorDisassembler to store the resulting dataframe of the ml.feature.StandardScaler.fit method into a CSV.

The code looks roughly like this:

val disassembler = new org.apache.spark.ml.feature.VectorDisassembler()
val disassembledDF = disassembler.setInputCol("scaledFeatures").transform(df)
disassembledDF.show()

来源：https://stackoverflow.com/questions/44160768/spark-2-1-cannot-write-vector-field-on-csv