12.5. Functions

Plugin Implementation

The function framework is used to implement SQL functions. Presto includes a number of built-in functions. In order to implement new functions, you can write a plugin that returns one more more functions from getFunctions():

  1. public class ExampleFunctionsPlugin
  2.         implements Plugin
  3. {
  4.     @Override
  5.     public Set<Class<?>> getFunctions()
  6.     {
  7.         return ImmutableSet.<Class<?>>builder()
  8.                 .add(ExampleNullFunction.class)
  9.                 .add(IsNullFunction.class)
  10.                 .add(IsEqualOrNullFunction.class)
  11.                 .add(ExampleStringFunction.class)
  12.                 .add(ExampleAverageFunction.class)
  13.                 .build();
  14.     }
  15. }

Note that the ImmutableSet class is a utility class from Guava. The getFunctions() method contains all of the classes for the functions that we will implement below in this tutorial.

For a full example in the codebase, see either the presto-ml module for machine learning functions or the presto-teradata-functions module for Teradata-compatible functions, both in the root of the Presto source.

Scalar Function Implementation

The function framework uses annotations to indicate relevant information about functions, including name, description, return type and parameter types. Below is a sample function which implements is_null:

  1. public class ExampleNullFunction
  2. {
  3.     @ScalarFunction("is_null", calledOnNullInput = true)
  4.     @Description("Returns TRUE if the argument is NULL")
  5.     @SqlType(StandardTypes.BOOLEAN)
  6.     public static boolean isNull(@SqlNullable @SqlType(StandardTypes.VARCHAR) Slice string)
  7.     {
  8.         return (string == null);
  9.     }
  10. }

The function is_null takes a single VARCHAR argument and returns a BOOLEAN indicating if the argument was NULL. Note that the argument to the function is of type Slice. VARCHAR uses Slice, which is essentially a wrapper around byte[], rather than String for its native container type.

  • @SqlType:

    The @SqlType annotation is used to declare the return type and the argument types. Note that the return type and arguments of the Java code must match the native container types of the corresponding annotations.

  • @SqlNullable:

    The @SqlNullable annotation indicates that the argument may be NULL. Without this annotation the framework assumes that all functions return NULL if any of their arguments are NULL. When working with a Type that has a primitive native container type, such as BigintType, use the object wrapper for the native container type when using @SqlNullable. The method must be annotated with @SqlNullable if it can return NULL when the arguments are non-null.

Parametric Scalar Functions

Scalar functions that have type parameters have some additional complexity. To make our previous example work with any type we need the following:

  1. @ScalarFunction(name = "is_null", calledOnNullInput = true)
  2. @Description("Returns TRUE if the argument is NULL")
  3. public final class IsNullFunction
  4. {
  5.     @TypeParameter("T")
  6.     @SqlType(StandardTypes.BOOLEAN)
  7.     public static boolean isNullSlice(@SqlNullable @SqlType("T") Slice value)
  8.     {
  9.         return (value == null);
  10.     }
  12.     @TypeParameter("T")
  13.     @SqlType(StandardTypes.BOOLEAN)
  14.     public static boolean isNullLong(@SqlNullable @SqlType("T") Long value)
  15.     {
  16.         return (value == null);
  17.     }
  19.     @TypeParameter("T")
  20.     @SqlType(StandardTypes.BOOLEAN)
  21.     public static boolean isNullDouble(@SqlNullable @SqlType("T") Double value)
  22.     {
  23.         return (value == null);
  24.     }
  26.     // ...and so on for each native container type
  27. }

  • @TypeParameter:

    The @TypeParameter annotation is used to declare a type parameter which can be used in the argument types @SqlType annotation, or return type of the function. It can also be used to annotate a parameter of type Type. At runtime, the engine will bind the concrete type to this parameter. @OperatorDependency may be used to declare that an additional function for operating on the given type parameter is needed. For example, the following function will only bind to types which have an equals function defined:

  1. @ScalarFunction(name = "is_equal_or_null", calledOnNullInput = true)
  2. @Description("Returns TRUE if arguments are equal or both NULL")
  3. public final class IsEqualOrNullFunction
  4. {
  5.     @TypeParameter("T")
  6.     @SqlType(StandardTypes.BOOLEAN)
  7.     public static boolean isEqualOrNullSlice(
  8.             @OperatorDependency(operator = OperatorType.EQUAL, returnType = StandardTypes.BOOLEAN, argumentTypes = {"T", "T"}) MethodHandle equals,
  9.             @SqlNullable @SqlType("T") Slice value1,
  10.             @SqlNullable @SqlType("T") Slice value2)
  11.     {
  12.         if (value1 == null && value2 == null) {
  13.             return true;
  14.         }
  15.         if (value1 == null || value2 == null) {
  16.             return false;
  17.         }
  18.         return (boolean) equals.invokeExact(value1, value2);
  19.     }
  21.     // ...and so on for each native container type
  22. }

Another Scalar Function Example

The lowercaser function takes a single VARCHAR argument and returns a VARCHAR, which is the argument converted to lower case:

  1. public class ExampleStringFunction
  2. {
  3.     @ScalarFunction("lowercaser")
  4.     @Description("converts the string to alternating case")
  5.     @SqlType(StandardTypes.VARCHAR)
  6.     public static Slice lowercaser(@SqlType(StandardTypes.VARCHAR) Slice slice)
  7.     {
  8.         String argument = slice.toStringUtf8();
  9.         return Slices.utf8Slice(argument.toLowerCase());
  10.     }
  11. }

Note that for most common string functions, including converting a string to lower case, the Slice library also provides implementations that work directly on the underlying byte[], which have much better performance. This function has no @SqlNullable annotations, meaning that if the argument is NULL, the result will automatically be NULL (the function will not be called).

Aggregation Function Implementation

Aggregation functions use a similar framework to scalar functions, but are a bit more complex.

  • AccumulatorState:

    All aggregation functions accumulate input rows into a state object; this object must implement AccumulatorState. For simple aggregations, just extend AccumulatorState into a new interface with the getters and setters you want, and the framework will generate all the implementations and serializers for you. If you need a more complex state object, you will need to implement AccumulatorStateFactory and AccumulatorStateSerializer and provide these via the AccumulatorStateMetadata annotation.

The following code implements the aggregation function avg_double which computes the average of a DOUBLE column:

  1. @AggregationFunction("avg_double")
  2. public class AverageAggregation
  3. {
  4.     @InputFunction
  5.     public static void input(LongAndDoubleState state, @SqlType(StandardTypes.DOUBLE) double value)
  6.     {
  7.         state.setLong(state.getLong() + 1);
  8.         state.setDouble(state.getDouble() + value);
  9.     }
  11.     @CombineFunction
  12.     public static void combine(LongAndDoubleState state, LongAndDoubleState otherState)
  13.     {
  14.         state.setLong(state.getLong() + otherState.getLong());
  15.         state.setDouble(state.getDouble() + otherState.getDouble());
  16.     }
  18.     @OutputFunction(StandardTypes.DOUBLE)
  19.     public static void output(LongAndDoubleState state, BlockBuilder out)
  20.     {
  21.         long count = state.getLong();
  22.         if (count == 0) {
  23.             out.appendNull();
  24.         }
  25.         else {
  26.             double value = state.getDouble();
  27.             DOUBLE.writeDouble(out, value / count);
  28.         }
  29.     }
  30. }

The average has two parts: the sum of the DOUBLE in each row of the column and the LONG count of the number of rows seen. LongAndDoubleState is an interface which extends AccumulatorState:

  1. public interface LongAndDoubleState
  2.         extends AccumulatorState
  3. {
  4.     long getLong();
  6.     void setLong(long value);
  8.     double getDouble();
  10.     void setDouble(double value);
  11. }

As stated above, for simple AccumulatorState objects, it is sufficient to just to define the interface with the getters and setters, and the framework will generate the implementation for you.

An in-depth look at the various annotations relevant to writing an aggregation function follows:

  • @InputFunction:

    The @InputFunction annotation declares the function which accepts input rows and stores them in the AccumulatorState. Similar to scalar functions you must annotate the arguments with @SqlType. Note that, unlike in the above scalar example where Slice is used to hold VARCHAR, the primitive double type is used for the argument to input. In this example, the input function simply keeps track of the running count of rows (via setLong()) and the running sum (via setDouble()).

  • @CombineFunction:

    The @CombineFunction annotation declares the function used to combine two state objects. This function is used to merge all the partial aggregation states. It takes two state objects, and merges the results into the first one (in the above example, just by adding them together).

  • @OutputFunction:

    The @OutputFunction is the last function called when computing an aggregation. It takes the final state object (the result of merging all partial states) and writes the result to a BlockBuilder.

  • Where does serialization happen, and what is GroupedAccumulatorState?

    The @InputFunction is usually run on a different worker from the @CombineFunction, so the state objects are serialized and transported between these workers by the aggregation framework. GroupedAccumulatorState is used when performing a GROUP BY aggregation, and an implementation will be automatically generated for you, if you don’t specify a AccumulatorStateFactory