scalaapache-sparkapache-spark-sqlspark-csv

inferSchema in spark-csv package

When CSV is read as dataframe in spark, all the columns are read as string. Is there any way to get the actual type of column?

I have the following csv file

Name,Department,years_of_experience,DOB
Sam,Software,5,1990-10-10
Alex,Data Analytics,3,1992-10-10

I've read the CSV using the below code

val df = sqlContext.
                  read.
                  format("com.databricks.spark.csv").
                  option("header", "true").
                  option("inferSchema", "true").
                  load(sampleAdDataS3Location)
df.schema

All the columns are read as string. I expect the column years_of_experience to be read as int and DOB to be read as date

Please note that I've set the option inferSchema to true.

I am using the latest version (1.0.3) of spark-csv package

Am I missing something here?

Solution

  • 2015-07-30

    The latest version is actually 1.1.0, but it doesn't really matter since it looks like inferSchema is not included in the latest release.

    2015-08-17

    The latest version of the package is now 1.2.0 (published on 2015-08-06) and schema inference works as expected:

    scala> df.printSchema
root
 |-- Name: string (nullable = true)
 |-- Department: string (nullable = true)
 |-- years_of_experience: integer (nullable = true)
 |-- DOB: string (nullable = true)

    Regarding automatic date parsing I doubt it will ever happen, or at least not without providing additional metadata.

    Even if all fields follow some date-like format it is impossible to say if a given field should be interpreted as a date. So it is either lack of out automatic date inference or spreadsheet like mess. Not to mention issues with timezones for example.

    Finally you can easily parse date string manually:

    sqlContext
  .sql("SELECT *, DATE(dob) as dob_d  FROM df")
  .drop("DOB")
  .printSchema

root
 |-- Name: string (nullable = true)
 |-- Department: string (nullable = true)
 |-- years_of_experience: integer (nullable = true)
 |-- dob_d: date (nullable = true)

    so it is really not a serious issue.

    2017-12-20:

    Built-in csv parser available since Spark 2.0 supports schema inference for dates and timestamp - it uses two options:

    See also How to force inferSchema for CSV to consider integers as dates (with "dateFormat" option)?