Minor Description
CSV files can have arbitrary column names:
scala> spark.range(1).select(col("id").as("Column?"), col("id")).write.option("header", true).csv("/tmp/foo") scala> spark.read.option("header", true).csv("/tmp/foo").schema res1: org.apache.spark.sql.types.StructType = StructType(StructField(Column?,StringType,true), StructField(id,StringType,true))
However, once a column with characters like "?" in the name gets used in a partitioning column, the column name gets silently escaped, and reading the schema information back renders the column name with "?" turned into "%3F":
scala> spark.range(1).select(col("id").as("Column?"), col("id")).write.partitionBy("Column?").option("header", true).csv("/tmp/bar") scala> spark.read.option("header", true).csv("/tmp/bar").schema res3: org.apache.spark.sql.types.StructType = StructType(StructField(id,StringType,true), StructField(Column%3F,IntegerType,true))
The same happens for other formats, but I encountered it working with CSV, since these more often contain ugly schemas...
Not sure if it's a bug or a feature, but it might be more intuitive to fail queries with invalid characters in the partitioning column name, rather than silently escaping the name?
