Interface for a function that produces a result value for each input row.

To evaluate each input row, Spark will first try to lookup and use a "magic method" (described below) through Java reflection. If the method is not found, Spark will call #produceResult(InternalRow) as a fallback approach.

The JVM type of result values produced by this function must be the type used by Spark's InternalRow API for the SQL data type returned by #resultType(). The mapping between DataType and the corresponding JVM type is defined below.

Magic method

IMPORTANT: the default implementation of #produceResult throws UnsupportedOperationException. Users must choose to either override this method, or implement a magic method with name #MAGIC_METHOD_NAME, which takes individual parameters instead of a InternalRow. The magic method approach is generally recommended because it provides better performance over the default #produceResult, due to optimizations such as whole-stage codegen, elimination of Java boxing, etc.

The type parameters for the magic method must match those returned from BoundFunction#inputTypes(). Otherwise Spark will not be able to find the magic method.

In addition, for stateless Java functions, users can optionally define the #MAGIC_METHOD_NAME as a static method, which further avoids certain runtime costs such as Java dynamic dispatch.

For example, a scalar UDF for adding two integers can be defined as follow with the magic method approach:

  public class IntegerAdd implementsScalarFunction {
    public DataType[] inputTypes() {
      return new DataType[] { DataTypes.IntegerType, DataTypes.IntegerType };
    }
    public int invoke(int left, int right) {
      return left + right;
    }
  }

In the above, since #MAGIC_METHOD_NAME is defined, and also that it has matching parameter types and return type, Spark will use it to evaluate inputs.

As another example, in the following:

  public class IntegerAdd implementsScalarFunction {
    public DataType[] inputTypes() {
      return new DataType[] { DataTypes.IntegerType, DataTypes.IntegerType };
    }
    public static int invoke(int left, int right) {
      return left + right;
    }
    public Integer produceResult(InternalRow input) {
      return input.getInt(0) + input.getInt(1);
    }
  }

the class defines both the magic method and the #produceResult, and Spark will use #MAGIC_METHOD_NAME over the #produceResult(InternalRow) as it takes higher precedence. Also note that the magic method is annotated as a static method in this case.

Resolution on magic method is done during query analysis, where Spark looks up the magic method by first converting the actual input SQL data types to their corresponding Java types following the mapping defined below, and then checking if there is a matching method from all the declared methods in the UDF class, using method name and the Java types.

Handling of nullable primitive arguments

The handling of null primitive arguments is different between the magic method approach and the #produceResult approach. With the former, whenever any of the method arguments meet the following conditions:

• the argument is of primitive type
• the argument is nullable
• the value of the argument is null

Spark will return null directly instead of calling the magic method. On the other hand, Spark will pass null primitive arguments to #produceResult and it is user's responsibility to handle them in the function implementation.

Because of the difference, if Spark users want to implement special handling of nulls for nullable primitive arguments, they should override the #produceResult method instead of using the magic method approach.

Spark data type to Java type mapping

The following are the mapping from SQL data type to Java type which is used by Spark to infer parameter types for the magic methods as well as return value type for #produceResult:

• org.apache.spark.sql.types.BooleanType: boolean
• org.apache.spark.sql.types.ByteType: byte
• org.apache.spark.sql.types.ShortType: short
• org.apache.spark.sql.types.IntegerType: int
• org.apache.spark.sql.types.LongType: long
• org.apache.spark.sql.types.FloatType: float
• org.apache.spark.sql.types.DoubleType: double
• org.apache.spark.sql.types.StringType: org.apache.spark.unsafe.types.UTF8String
• org.apache.spark.sql.types.DateType: int
• org.apache.spark.sql.types.TimestampType: long
• org.apache.spark.sql.types.BinaryType: byte[]
• org.apache.spark.sql.types.DayTimeIntervalType: long
• org.apache.spark.sql.types.YearMonthIntervalType: int
• org.apache.spark.sql.types.DecimalType: org.apache.spark.sql.types.Decimal
• org.apache.spark.sql.types.StructType: InternalRow
• org.apache.spark.sql.types.ArrayType: org.apache.spark.sql.catalyst.util.ArrayData
• org.apache.spark.sql.types.MapType: org.apache.spark.sql.catalyst.util.MapData