How to add a scalar function?

Simple Functions

This document describes the main concepts, features, and examples of the simple function API in Velox. For more real-world API usage examples, check velox/example/SimpleFunctions.cpp.

A simple scalar function, e.g. a mathematical function, can be added by wrapping a C++ function in a templated class. For example, a ceil function can be implemented as:

template <typename TExec>
struct CeilFunction {
  VELOX_DEFINE_FUNCTION_TYPES(TExec);

  template <typename T>
  FOLLY_ALWAYS_INLINE void call(T& result, const T& a) {
    result = std::ceil(a);
  }
};

All simple function classes need to be templated, and provide a “call” method (or one of the variations described below). The top-level template parameter provides the type system adapter, which allows developers to use non-primitive types such as strings, arrays, maps, and struct (check below for examples). Although the top-level template parameter is not used for functions operating on primitive types, such as the one in the example above, it still needs to be specified.

The call method itself can also be templated or overloaded to allow the function to be called on different input types, e.g. float and double. Note that template instantiation will only happen during function registration, described in the “Registration” section below.

Do not use legacy VELOX_UDF_BEGIN and VELOX_UDF_END macros.

The “call” function (or one of its variations) may return (a) void indicating the function never returns null values, or (b) boolean indicating whether the result of the computation is null. The meaning of the returned boolean is “result was set”, i.e. true means non-null result was populated, false means no (null) result. If “ceil(0)” were to return a null, the function could be re-written as follows:

template <typename TExec>
struct NullableCeilFunction {
  VELOX_DEFINE_FUNCTION_TYPES(TExec);

  template <typename T>
  FOLLY_ALWAYS_INLINE bool call(T& result, const T& a) {
    result = std::ceil(a);
    return a != 0; // Return NULL if input is zero.
  }
};

The argument list must start with an output parameter “result” followed by the function arguments. The “result” argument must be a reference. Function arguments must be const references. The C++ types of the function arguments and the result argument must match Velox types.

Velox Type	C++ Argument Type	C++ Result Type
BOOLEAN	arg_type<bool>	out_type<bool>
TINYINT	arg_type<int8_t>	out_type<int8_t>
SMALLINT	arg_type<int16_t>	out_type<int16_t>
INTEGER	arg_type<int32_t>	out_type<int32_t>
BIGINT	arg_type<int64_t>	out_type<int64_t>
REAL	arg_type<float>	out_type<float>
DOUBLE	arg_type<double>	out_type<double>
TIMESTAMP	arg_type<Timestamp>	out_type<Timestamp>
DATE	arg_type<Date>	out_type<Date>
VARCHAR	arg_type<Varchar>	out_type<Varchar>
VARBINARY	arg_type<Varbinary>	out_type<Varbinary>
ARRAY	arg_type<Array<E>>	out_type<Array<E>>
MAP	arg_type<Map<K,V>>	out_type<Map<K, V>>
ROW	arg_type<Row<T1, T2, T3,…>>	out_type<Row<T1, T2, T3,…>>

arg_type and out_type templates are defined by the VELOX_DEFINE_FUNCTION_TYPES(TExec) macro in the struct definition. For primitive types, arg_type<T> is the same as out_type<T> and the same as T. This holds for boolean, integers, floating point types and timestamp. For DATE, arg_type<Date> is the same as out_type<Date> and is defined as int32_t.

A signature of a function that takes an integer and a double and returns a double would look like this:

void call(arg_type<double>& result, const arg_type<int32_t>& a, const arg_type<double>& b)

Which is equivalent to

void call(double& result, const int32_t& a, const double& b)

For strings, arg_type<Varchar> is defined as StringView, while out_type<Varchar> is defined as StringWriter.

arg_type and out_type for Varchar, Array, Map and Row provide interfaces similar to std::string, std::vector, std::unordered_map and std::tuple. The underlying implementations are optimized to read and write from and to the columnar representation without extra copying. More explanation and the APIs of the arg_type and out_type for string and complex types can be found in View and Writer Types.

Note: Do not pay too much attention to complex type mappings at the moment. They are included here for completeness.

Null Behavior

Most functions have default null behavior, e.g. a null value in any of the arguments produces a null result. The expression evaluation engine automatically produces nulls for such inputs, eliding a call to the actual function. If a given function has a different behavior for null inputs, it must define a “callNullable” function instead of a “call” function. Here is an artificial example of a ceil function that returns 0 for null input:

template <typename TExec>
struct CeilFunction {
  template <typename T>
  FOLLY_ALWAYS_INLINE void callNullable(T& result, const T* a) {
    // Return 0 if input is null.
    if (a) {
      result = std::ceil(*a);
    } else {
      result = 0;
    }
  }
};

Notice that callNullable function takes arguments as raw pointers and not references to allow for specifying null values. callNullable() can also return void to indicate that the function does not produce null values.

Null-Free Fast Path

A “callNullFree” function may be implemented in place of or along side “call” and/or “callNullable” functions. When only the “callNullFree” function is implemented, evaluation of the function will be skipped and null will automatically be produced if any of the input arguments are null (like default null behavior) or if any of the input arguments are of a complex type and contain null anywhere in their value, e.g. an array that has a null element. If “callNullFree” is implemented alongside “call” and/or “callNullable”, an O(N * D) check is applied to the batch to see if any of the input arguments may be or contain null, where N is the number of input arguments and D is the depth of nesting in complex types. Only if it can definitively be determined that there are no nulls will “callNullFree” be invoked. In this case, “callNullFree” can act as a fast path by avoiding any per row null checks.

Here is an example of an array_min function that returns the minimum value in an array:

template <typename TExec>
struct ArrayMinFunction {
  VELOX_DEFINE_FUNCTION_TYPES(TExec);

  template <typename TInput>
  FOLLY_ALWAYS_INLINE bool callNullFree(
      TInput& out,
      const null_free_arg_type<Array<TInput>>& array) {
    out = INT32_MAX;
    for (auto i = 0; i < array.size(); i++) {
      if (array[i] < out) {
        out = array[i]
      }
    }
    return true;
  }
};

Notice that we can access the elements of “array” without checking their nullity in “callNullFree”. Also notice that we wrap the input type in the null_free_arg_type<…> template instead of the arg_type<…> template. This is required as the input types for complex types are of a different type in “callNullFree” functions that do not wrap values in an std::optional-like interface upon access.

Determinism

By default simple functions are assumed to be deterministic, e.g. given the same inputs they always produce the same results. If this is not the case, the function must define a static constexpr bool is_deterministic member:

static constexpr bool is_deterministic = false;

An example of such function is rand():

template <typename TExec>
struct RandFunction {
  static constexpr bool is_deterministic = false;

  FOLLY_ALWAYS_INLINE bool call(double& result) {
    result = folly::Random::randDouble01();
    return true;
  }
};

All-ASCII Fast Path

Functions that process string inputs must work correctly for UTF-8 inputs. However, these functions often can be implemented more efficiently if input is known to contain only ASCII characters. Such functions can provide a “call” method to process UTF-8 strings and a “callAscii” method to process ASCII-only strings. The engine will check the input strings and invoke “callAscii” method if input is all ASCII or “call” if input may contain multi-byte characters.

In addition, most functions that take string inputs and produce a string output have so-called default ASCII behavior, e.g. all-ASCII input guarantees all-ASCII output. If that’s the case, the function can indicate so by defining the is_default_ascii_behavior member variable and initializing it to true. The engine will automatically mark the result strings as all-ASCII. When these strings are passed as input to some other function, the engine won’t need to scan the strings to determine whether they are ASCII or not.

Here is an example of a trim function:

template <typename TExec>
struct TrimFunction {
  VELOX_DEFINE_FUNCTION_TYPES(TExec);

  // ASCII input always produces ASCII result.
  static constexpr bool is_default_ascii_behavior = true;

  // Properly handles multi-byte characters.
  FOLLY_ALWAYS_INLINE bool call(
      out_type<Varchar>& result,
      const arg_type<Varchar>& input) {
    stringImpl::trimUnicodeWhiteSpace<leftTrim, rightTrim>(result, input);
    return true;
  }

  // Assumes input is all ASCII.
  FOLLY_ALWAYS_INLINE bool callAscii(
      out_type<Varchar>& result,
      const arg_type<Varchar>& input) {
    stringImpl::trimAsciiWhiteSpace<leftTrim, rightTrim>(result, input);
    return true;
  }
};

Zero-copy String Result

Functions like substr() and trim() can produce zero-copy results by referencing input strings. To do that they must define a reuse_strings_from_arg member variable and initialize it to the index of the argument whose strings are being re-used in the result. This will allow the engine to add a reference to input string buffers to the result vector and ensure that these buffers will not go away prematurely. The output types can be scalar strings (varchar and varbinaries), but also complex types containing strings, such as arrays, maps, and rows.

The setNoCopy method of the out_type template can be used to set the result to a string in the input argument without copying. The setEmpty method can be used to set the result to an empty string.

// Results refer to strings in the first argument.
static constexpr int32_t reuse_strings_from_arg = 0;

Here is an example of a zero-copy function:

template <typename TExec>
struct TrimFunction {
  VELOX_DEFINE_FUNCTION_TYPES(TExec);

  // Results refer to strings in the first argument.
  static constexpr int32_t reuse_strings_from_arg = 0;

  FOLLY_ALWAYS_INLINE void call(
      out_type<Varchar>& result,
      const arg_type<Varchar>& input) {
    if (input.size() == 0) {
      result.setEmpty();
      return;
    }
    result.setNoCopy(stringImpl::trimUnicodeWhiteSpace(input));
  }
};

Access to Session Properties and Constant Inputs

Some functions require access to session properties such as session’s timezone. Some examples are the day(), hour(), and minute() Presto functions. Other functions could benefit from pre-processing some of the constant inputs, e.g. compile regular expression patterns or parse date and time units. To get access to session properties and constant inputs the function must define an initialize method which receives a constant reference to QueryConfig and a list of constant pointers for each of the input arguments. Constant inputs will have their values specified. Inputs which are not constant will be passed as nullptr’s. The signature of the initialize method is similar to that of callNullable method with an additional first parameter const core::QueryConfig&. The engine calls the initialize method once per query and thread of execution.

Here is an example of an hour function extracting time zone from the session properties and using it when processing inputs.

template <typename TExec>
struct HourFunction {
  VELOX_DEFINE_FUNCTION_TYPES(TExec);

  const tz::TimeZone* timeZone_ = nullptr;

  FOLLY_ALWAYS_INLINE void initialize(
      const std::vector<TypePtr>& inputTypes,
      const core::QueryConfig& config,
      const arg_type<Timestamp>* /*timestamp*/) {
    timeZone_ = getTimeZoneFromConfig(config);
  }

  FOLLY_ALWAYS_INLINE bool call(
      int64_t& result,
      const arg_type<Timestamp>& timestamp) {
    int64_t seconds = getSeconds(timestamp, timeZone_);
    std::tm dateTime;
    gmtime_r((const time_t*)&seconds, &dateTime);
    result = dateTime.tm_hour;
    return true;
  }
};

Here is another example of the date_trunc() function parsing the constant unit argument during initialize and re-using parsed value when processing individual rows.

template <typename TExec>
struct DateTruncFunction {
  VELOX_DEFINE_FUNCTION_TYPES(TExec);

  const tz::TimeZone* timeZone_ = nullptr;
  std::optional<DateTimeUnit> unit_;

  FOLLY_ALWAYS_INLINE void initialize(
      const std::vector<TypePtr>& inputTypes,
      const core::QueryConfig& config,
      const arg_type<Varchar>* unitString,
      const arg_type<Timestamp>* /*timestamp*/) {
    timeZone_ = getTimeZoneFromConfig(config);
    if (unitString != nullptr) {
      unit_ = fromDateTimeUnitString(*unitString);
    }
  }

  FOLLY_ALWAYS_INLINE bool call(
      out_type<Timestamp>& result,
      const arg_type<Varchar>& unitString,
      const arg_type<Timestamp>& timestamp) {
    const auto unit =
        unit_.has_value() ? unit_.value() : fromDateTimeUnitString(unitString);
    ...<use unit enum>...
  }
};

If the initialize() method throws, the exception will be captured and reported as output for every single active row. If there are no active rows, the exception will not be raised.

Registration

Use registerFunction template to register simple functions.

template <template <class> typename Func, typename TReturn, typename... TArgs>
void registerFunction(
    const std::vector<std::string>& aliases = {},
    std::shared_ptr<const Type> returnType = nullptr)

The first template parameter is the class name, the next template parameter is the return type, the remaining template parameters are argument types. Aliases parameter allows developers to specify multiple names for the same function, but each function registration needs to provide at least one name. The “ceil” function defined above can be registered using the following function call:

registerFunction<CeilFunction, double, double>({"ceil", "ceiling");

Here, we register the CeilFunction function that takes a double and returns a double. If we want to allow the ceil function to be called on float inputs, we need to call registerFunction again:

registerFunction<CeilFunction, float, float>({"ceil", "ceiling");

We need to call registerFunction for each signature we want to support.

Here is a mapping from Velox types to C++ types that should be used for argument and return types during registration.

Velox Type	C++ Type
BOOLEAN	bool
TINYINT	int8_t
SMALLINT	int16_t
INTEGER	int32_t
BIGINT	int64_t
REAL	float
DOUBLE	double
TIMESTAMP	Timestamp
DATE	Date
VARCHAR	Varchar
VARBINARY	Varbinary
ARRAY	Array<E>
MAP	Map<K,V>
ROW	Row<T1, T2, T3,…>

For example, to register array_min function for string inputs:

registerFunction<ArrayMinFunction, Varchar, Array<Varchar>>({"array_min"});

To register array_min function for arrays of any type, use Generic<T1> for the element type:

registerFunction<ArrayMinFunction, Generic<T1>, Array<Generic<T1>>>({"array_min"});

Since array_min needs to sort the elements to find the smallest, the element type needs to be orderable. You can restrict array elements to orderable types using Orderable<T1>.

registerFunction<ArrayMinFunction, Orderable<T1>, Array<Orderable<T1>>>({"array_min"});

You can use multiple generic types in a function signature. For example, to register map_top_n function:

registerFunction<
    MapTopNFunction,
    Map<Generic<T1>, Orderable<T2>>,    // result map type
    Map<Generic<T1>, Orderable<T2>>,    // input map type
    int64_t                             // type of N argument
>({"map_top_n"});

Generic types must use T1, T2, T3… naming.

Finally, you can specify that an argument must be constant using Constant<T>. For example, to specify rand signature with a constant seed argument:

registerFunction<RandFunction, double, Constant<int32_t>>({"rand"});

Variadic Arguments

The last argument to a simple function may be marked “Variadic”. This means invocations of this function may include 0..N arguments of that type at the end of the call. While not a true type in Velox, “Variadic” can be thought of as a syntactic type, and behaves somewhat similarly to Array.

C++ Argument Type	C++ Actual Argument Type
arg_type<Variadic<E>>	NullableVariadicView<E>
null_free_arg_type<Variadic<E>>	NullFreeVariadicView<E>

Like the NullableArrayView and NullFreeArrayView, VariadicViews has a similar interface to const std::vector<std::optional<V>>.

NullableVariadicView, and NullFreeVariadicView, supports the following:

• size_t size() : return the number of arguments that were passed as part of the “Variadic” type in the function invocation.

• operator[](size_t index) : access the value of the argument at index. It returns either null_free_arg_type<E> or OptionalAccessor<E>.

• VariadicView<T>::Iterator begin() : iterator to the first argument.

• VariadicView<T>::Iterator end() : iterator indicating end of iteration.

• bool mayHaveNulls() : a check on the nullity of the arugments (note this takes time proportional to the number of arguments). When it returns false, there are definitely no nulls, a true does not guarantee null existence.

• VariadicView<T>::SkipNullsContainer SkipNulls() : return an iterable container that provides direct access to each argument with a non-null value.

The code below shows an example of a function that concatenates a variable number of strings:

template <typename T>
struct VariadicArgsReaderFunction {
  VELOX_DEFINE_FUNCTION_TYPES(T);

  FOLLY_ALWAYS_INLINE bool call(
      out_type<Varchar>& out,
      const arg_type<Variadic<Varchar>>& inputs) {
    for (const auto& input : inputs) {
      if (input.has_value()) {
        output += input.value();
      }
    }

    return true;
  }
};

Vector Functions

Simple functions process a single row and produce a single value as a result. Vector functions process a batch or rows and produce a vector of results. When implementing a function, simple function is preferred unless the implementation of vector function provides a significant performance gain which can be demonstrated with a benchmark. Some of the defining features of vector functions are:

• take vectors as inputs and produce vectors as a result;

• have access to vector encodings and metadata;

• can be defined for generic input types, e.g. generic arrays, maps and structs;

• allow for implementing lambda functions;

Vector function interface allows for many optimizations that are not available to simple functions. These optimizations often leverage different vector encodings and columnar representations of the vectors. Here are some examples,

• map_keys() function takes advantage of the ArrayVector representation and simply returns the inner “keys” vector without doing any computation. Similarly, map_values() function simply returns the inner “values” vector.

• map_entries() function takes the pieces of the input vector - “nulls”, “sizes” and “offsets” buffers and “keys” and “values” vectors - and simply repackages them in the form of a RowVector.

• cardinality() function takes advantage of the ArrayVector and MapVector representations and simply returns the “sizes” buffer of the input vector.

• is_null() function copies the “nulls” buffer of the input vector, flips the bits in bulk and returns the result.

• element_at() function and subscript operator for arrays and maps use dictionary encoding to represent a subset of the input “elements” or “values” vector without copying.

To define a vector function, make a subclass of exec::VectorFunction and implement the “apply” method.

void apply(
      const SelectivityVector& rows,
      std::vector<VectorPtr>& args,
      Expr* caller,
      EvalCtx& context,
      VectorPtr& result) const

Input rows

The “rows” parameter specifies the set of rows in the incoming batch to process. This set may not include all the rows. By default, a vector function is assumed to have the default null behavior, e.g. null in any input produces a null result. In this case, the expression evaluation engine will exclude rows with nulls from the “rows” specified in the call to “apply”. If a function has a different behavior for null inputs, it must specify that during registration. See vector function registration for more details.

In this case, the “rows” parameter will include rows with null inputs and the function will need to handle these. By default, the function can assume that all inputs are not null for all “rows”.

When evaluating a function as part of a conditional expression, e.g. AND, OR, IF, SWITCH, the set of “rows” represents a subset of the rows that need evaluating. Consider some examples.

a > 5 AND b > 7

Here, a > 5 is evaluated on all rows where “a” is not null, but b > 7 is evaluated on rows where b is not null and a > 5 is true.

IF(condition, a + 5, b - 3)

Here, a + 5 is evaluated on rows where a is not null and condition is true, while b - 3 is evaluated on rows where b is not null and condition is not true.

In some cases, the values outside of “rows” may be undefined, uninitialized or contain garbage. This would be the case if an earlier filter operation produced dictionary-encoded vectors with indices pointing to a subset of the rows which passed the filter. When evaluating f(g(a)), where a = Dict (a0), function “g” is evaluated on a subset of rows in “a0” and may produce a result where only that subset of rows is populated. Then, function “f” is evaluated on the same subset of rows in the result of “g”. The input to “f” will have values outside of “rows” undefined, uninitialized or contain garbage.

Note that SelectivityVector::applyToSelected method can be used to loop over the specified rows in a way that’s rather similar to a standard for loop.

rows.applyToSelected([&] (auto row) {
    // row is the 0-based row number
    // .... process the row
});

Input vectors

The “args” parameter is an std::vector of Velox vectors containing the values of the function arguments. These vectors are not necessarily flat and may be dictionary or constant encoded. However, a deterministic function that takes a single argument and has default null behavior is guaranteed to receive its only input as a flat or constant vector. By default, a function is assumed to be deterministic. If that’s not the case, it must specify the non-deterministic behavior during registration. See vector function registration for more details.

Note that DecodedVector can be used to get a flat vector-like interface to any vector. A helper class exec::DecodedArgs can be used to decode multiple arguments.

exec::DecodedArgs decodedArgs(rows, args, context);

auto firstArg = decodedArgs.at(0);
auto secondArg = decodedArgs.at(1);

Result vector

The “result” parameter is a raw pointer to VectorPtr, which is a std::shared_ptr to BaseVector. It can be null, may point to a scratch vector that is maybe reusable or a partially populated vector whose contents must be preserved.

A partially populated vector is specified when evaluating the “else” branch of an IF. In this case, the results of the “then” branch must be preserved. This can be easily achieved by following one of the two patterns.

Calculate the result for all or just the specified rows into a new vector, then use EvalCtx::moveOrCopyResult method to either std::move the vector into “result” or copy individual rows into partially populated “result”.

Here is an example of using moveOrCopyResult to implement map_keys function:

void apply(
    const SelectivityVector& rows,
    std::vector<VectorPtr>& args,
    exec::Expr* /* caller */,
    exec::EvalCtx& context,
    VectorPtr& result) const override {
  auto mapVector = args[0]->as<MapVector>();
  auto mapKeys = mapVector->mapKeys();

  auto localResult = std::make_shared<ArrayVector>(
      context.pool(),
      ARRAY(mapKeys->type()),
      mapVector->nulls(),
      rows.end(),
      mapVector->offsets(),
      mapVector->sizes(),
      mapKeys,
      mapVector->getNullCount());

  context.moveOrCopyResult(localResult, rows, result);
}

Use BaseVector::ensureWritable method to initialize “result” to a flat uniquely-referenced vector while preserving values in rows not specified in “rows”. Then, calculate and fill in the “rows” in “result”. BaseVector::ensureWritable creates a new vector if “result” is null. If result is not null, but not-flat or not singly-referenced, BaseVector::ensureWritable creates a new vector and copies non-”rows” values from “result” into the newly created vector. If “result” is not null and flat, BaseVector::ensureWritable checks the inner buffers and copies these if they are not singly referenced. BaseVector::ensureWritable also recursively calls itself on inner vectors (elements vector for the array, keys and values for map, fields for struct) to make sure the vector is “writable” all the way through.

Here is an example of using BaseVector::ensureWritable to implement cardinality function for maps:

void apply(
    const SelectivityVector& rows,
    std::vector<VectorPtr>& args,
    exec::Expr* /* caller */,
    exec::EvalCtx& context,
    VectorPtr& result) const override {

  BaseVector::ensureWritable(rows, BIGINT(), context.pool(), result);
  BufferPtr resultValues =
       result->as<FlatVector<int64_t>>()->mutableValues(rows.size());
  auto rawResult = resultValues->asMutable<int64_t>();

  auto mapVector = args[0]->as<MapVector>();
  auto rawSizes = mapVector->rawSizes();

  rows.applyToSelected([&](vector_size_t row) {
    rawResult[row] = rawSizes[row];
  });
}

Simple implementation

Vector function interface is very flexible and allows for many interesting optimizations. It may also feel very complicated. Let’s see how we can use DecodedVector and BaseVector::ensureWritable to implement the “power(a, b)” function as a vector function in a way that is not much more complicated than the simple function. To clarify, it is best to implement the “power” function as a simple function. I’m using it here for illustration purposes only.

// Initialize flat results vector.
BaseVector::ensureWritable(rows, DOUBLE(), context.pool(), result);
auto rawResults = result->as<FlatVector<int64_t>>()->mutableRawValues();

// Decode the arguments.
DecodedArgs decodedArgs(rows, args, context);
auto base = decodedArgs.at(0);
auto exp = decodedArgs.at(1);

// Loop over rows and calculate the results.
rows.applyToSelected([&](int row) {
  rawResults[row] =
      std::pow(base->valueAt<double>(row), exp->valueAt<double>(row));
});

You may want to optimize for the case when both base and exponent being flat and eliminate the overhead of calling DecodedVector::valueAt template.

if (base->isIdentityMapping() && exp->isIdentityMapping()) {
  auto baseValues = base->data<double>();
  auto expValues = exp->data<double>();
  rows.applyToSelected([&](int row) {
    rawResults[row] = std::pow(baseValues[row], expValues[row]);
  });
} else {
  rows.applyToSelected([&](int row) {
    rawResults[row] =
        std::pow(base->valueAt<double>(row), exp->valueAt<double>(row));
  });
}

You may decide to further optimize for the case of flat base and constant exponent.

if (base->isIdentityMapping() && exp->isIdentityMapping()) {
  auto baseValues = base->data<double>();
  auto expValues = exp->data<double>();
  rows.applyToSelected([&](int row) {
    rawResults[row] = std::pow(baseValues[row], expValues[row]);
  });
} else if (base->isIdentityMapping() && exp->isConstantMapping()) {
  auto baseValues = base->data<double>();
  auto expValue = exp->valueAt<double>(0);
  rows.applyToSelected([&](int row) {
    rawResults[row] = std::pow(baseValues[row], expValue);
  });
} else {
  rows.applyToSelected([&](int row) {
    rawResults[row] =
        std::pow(base->valueAt<double>(row), exp->valueAt<double>(row));
  });
}

Hopefully, you can see now that additional complexity in the implementation comes only from introducing optimization paths. Developers need to decide whether that complexity is justified on a case by case basis.

TRY expression support

A built-in TRY expression evaluates input expression and handles certain types of errors by returning NULL. It is used for the cases where it is preferable that queries produce NULL or default values instead of failing when corrupt or invalid data is encountered. To specify default values, the TRY expression can be used in conjunction with the COALESCE function.

The implementation of the TRY expression relies on the VectorFunction implementation to call EvalCtx::setError(row, exception) instead of throwing exceptions directly.

void setError(vector_size_t index, const std::exception_ptr& exceptionPtr);

A typical pattern would be to loop over rows, apply a function wrapped in a try-catch and call context->setError(row, std::current_exception()); from the catch block.

rows.applyToSelected([&](auto row) {
  try {
    // ... calculate and store the result for the row
  } catch (const std::exception& e) {
    context.setError(row, std::current_exception());
  }
});

There is an EvalCtx::applyToSelectedNoThrow convenience method that can be used instead of the explicit try-catch block above:

context.applyToSelectedNoThrow(rows, [&](auto row) {
  // ... calculate and store the result for the row
});

Simple functions are compatible with the TRY expression by default. The framework wraps the “call” and “callNullable” methods in a try-catch and reports errors using context.setError.

Registration

Use exec::registerVectorFunction to register a stateless vector function.

bool registerVectorFunction(
    const std::string& name,
    std::vector<FunctionSignaturePtr> signatures,
    std::unique_ptr<VectorFunction> func,
    VectorFunctionMetadata metadata = {},
    bool overwrite = true)

exec::registerVectorFunction takes a name, a list of supported signatures and unique_ptr to an instance of the function. It takes an optional ‘metadata’ parameter that specifies whether a function is deterministic, has default null behavior, and other properties. A helper VectorFunctionMetadataBuilder class allows to easily construct ‘metadata’. For example,

VectorFunctionMetadataBuilder().defaultNullBehavior(false).build();

An optional “overwrite” flag specifies whether to overwrite a function if a function with the specified name already exists.

Use exec::registerStatefulVectorFunction to register a stateful vector function.

Note: A vector function will be given precedence over a simple function during resolution time. This is because in certain cases it makes sense to write an optimized vector function, and thus more precedence is given to a vector function over an equivalent simple function.

bool registerStatefulVectorFunction(
    const std::string& name,
    std::vector<FunctionSignaturePtr> signatures,
    VectorFunctionFactory factory,
    VectorFunctionMetadata metadata = {},
    bool overwrite = true)

exec::registerStatefulVectorFunction takes a name, a list of supported signatures and a factory function that can be used to create an instance of the vector function. Expression evaluation engine uses a factory function to create a new instance of the vector function for each thread of execution. In a single-threaded execution, a single instance of the function is used to process all batches of data. In a multi-threaded execution, each thread makes a separate instance of the function.

Factory function is called with a function name, types and optionally constant values for the arguments. For example, regular expressions functions are often called with constant regular expressions. A stateful vector function can compile the regular expression once (per thread of execution) and reuse the compiled expression for multiple batches of data. Similarly, an IN expression used with a constant IN-list can create a hash set of the values once and reuse it for all the batches of data.

// Represents arguments for stateful vector functions. Stores element type, and
// the constant value (if supplied).
struct VectorFunctionArg {
  const TypePtr type;
  const VectorPtr constantValue;
};

using VectorFunctionFactory = std::function<std::shared_ptr<VectorFunction>(
    const std::string& name,
    const std::vector<VectorFunctionArg>& inputArgs)>;

Function signature

It is recommended to use FunctionSignatureBuilder to create FunctionSignature instances. FunctionSignatureBuilder and FunctionSignature support Java-like generics, variable number of arguments and lambdas. Here are some examples.

The length function takes a single argument of type varchar and returns a bigint:

// varchar -> bigint
exec::FunctionSignatureBuilder()
  .returnType("bigint")
  .argumentType("varchar")
  .build()

The substr function takes a varchar and two integers for start and length. To specify types of multiple arguments, call argumentType() method for each argument in order.

// varchar, integer, integer -> bigint
exec::FunctionSignatureBuilder()
  .returnType("varchar")
  .argumentType("varchar")
  .argumentType("integer")
  .argumentType("integer")
  .build()

The concat function takes an arbitrary number of varchar inputs and returns a varchar. FunctionSignatureBuilder allows specifying that the last augment may appear zero or more times by calling variableArity(“varchar”) method.

// varchar... -> varchar
exec::FunctionSignatureBuilder()
    .returnType("varchar")
    .variableArity("varchar")
    .build()

The map_keys function takes any map and returns an array of map keys.

// map(K,V) -> array(K)
exec::FunctionSignatureBuilder()
  .knownTypeVariable("K")
  .typeVariable("V")
  .returnType("array(K)")
  .argumentType("map(K,V)")
  .build()

The transform function takes an array and a lambda, applies the lambda to each element of the array and returns a new array of the results.

// array(T), function(T, U) -> array(U)
exec::FunctionSignatureBuilder()
  .typeVariable("T")
  .typeVariable("U")
  .returnType("array(U)")
  .argumentType("array(T)")
  .argumentType("function(T, U)")
  .build();

The signature of a function that handles DECIMAL types can additionally take variables and constraints to represent the precision and scale values. The constraints are evaluated using a type calculator built from Flex and Bison tools. The decimal arithmetic addition function has the following signature:

// decimal, decimal -> decimal
exec::FunctionSignatureBuilder()
  .returnType("DECIMAL(r_precision, r_scale)")
  .argumentType("DECIMAL(a_precision, a_scale)")
  .argumentType("DECIMAL(b_precision, b_scale)")
  .variableConstraint(
      "r_precision",
      "min(38, max(a_precision - a_scale, b_precision - b_scale) + max(a_scale, b_scale) + 1)")
  .variableConstraint("r_scale", "max(a_scale, b_scale)")
  .build();

The type names used in FunctionSignatureBuilder can be either lowercase standard types, a special type “any”, or the ones defined by calling typeVariable() method. “any” type can be used to specify a printf-like function which takes any number of arguments of any possibly non-matching types.

Testing

Add a test using FunctionBaseTest from velox/functions/prestosql/tests/utils/FunctionBaseTest.h as a base class. Name your test and the .cpp file <function-name>Test, e.g. CardinalityTest in CardinalityTest.cpp or IsNullTest in IsNullTest.cpp.

FunctionBaseTest has many helper methods for generating test vectors. It also provides an evaluate() method that takes a SQL expression and input data, evaluates the expression and returns the result vector. SQL expression is parsed using DuckDB and type resolution logic is leveraging the function signatures specified during registration. assertEqualVectors() method takes two vectors, expected and actual, and asserts that they represent the same values. The encodings of the vectors may not be the same.

Here is an example of a test for vector function “contains”:

TEST_F(ArrayContainsTest, integerWithNulls) {
  auto arrayVector = makeNullableArrayVector<int64_t>(
      {{1, 2, 3, 4},
       {3, 4, 5},
       {},
       {5, 6, std::nullopt, 7, 8, 9},
       {7, std::nullopt},
       {10, 9, 8, 7}});

  auto testContains = [&](std::optional<int64_t> search,
                          const std::vector<std::optional<bool>>& expected) {
    auto result = evaluate<SimpleVector<bool>>(
        "contains(c0, c1)",
        makeRowVector({
            arrayVector,
            makeConstant(search, arrayVector->size()),
        }));

    assertEqualVectors(makeNullableFlatVector<bool>(expected), result);
  };

  testContains(1, {true, false, false, std::nullopt, std::nullopt, false});
  testContains(3, {true, true, false, std::nullopt, std::nullopt, false});
  testContains(5, {false, true, false, true, std::nullopt, false});
  testContains(7, {false, false, false, true, true, true});
  testContains(-2, {false, false, false, std::nullopt, std::nullopt, false});
}

Tests for simple functions could benefit from using the evaluateOnce () template which takes SQL expression and scalar values for the inputs, evaluates the expression on a vector of length 1 and returns the scalar result. Here is an example of a test for simple function “sqrt”:

TEST_F(ArithmeticTest, sqrt) {
  constexpr double kDoubleMax = std::numeric_limits<double>::max();
  const double kNan = std::numeric_limits<double>::quiet_NaN();

  const auto sqrt = [&](std::optional<double> a) {
    return evaluateOnce<double>("sqrt(c0)", a);
  };

  EXPECT_EQ(1.0, sqrt(1));
  EXPECT_THAT(sqrt(-1.0), IsNan());
  EXPECT_EQ(0, sqrt(0));

  EXPECT_EQ(2, sqrt(4));
  EXPECT_EQ(3, sqrt(9));
  EXPECT_FLOAT_EQ(1.34078e+154, sqrt(kDoubleMax).value_or(-1));
  EXPECT_EQ(std::nullopt, sqrt(std::nullopt));
  EXPECT_THAT(sqrt(kNan), IsNan());
}

Function names

For both simple and vector functions, their names are case insensitive. Function names are converted to lower case automatically when the functions are registered and when they are resolved for a given expression.

The following names are reserved for special forms and cannot be used as function names:

• and

• or

• cast

• if

• switch

• coalesce

• try

• row_constructor

Function Resolution order

Vector functions have precedence over simple functions during function resolution. If a function foo has multiple implementations, then the order in which function resolution will proceed is as follows:

1. Vector Function

2. Simple Function which are generic free and variadic free

3. Simple Function has variadic but generic free

4. Simple Function has generic but no variadic of generic

5. Simple function has variadic of generic

The available function with lowest rank is picked during function resolution. If there is more than one function with the same lowest rank, we count the number of concrete types in the signature and return the signature with highest concrete types count. (a concrete type is any type other than variadic or generic).

For example: consider the two signatures bellow which are both of type 4.

void call(bool& out, const int& , const Any& , const& Variadic<int>)    // concrete types = 2
void call(bool& out, const int& , const Any& ,const Any&)               // concrete types = 1

When both of them are valid for a given input, the first one will be picked since it has more concrete types. When number of concrete types are the same, the call is ambiguous, and it’s undefined which function is called.

Benchmarking

Add a benchmark using folly::Benchmark framework and FunctionBenchmarkBase from velox/functions/lib/benchmarks/FunctionBenchmarkBase.h as a base class. Benchmarks are a great way to check if an optimization is working, evaluate how much benefit it brings and decide whether it is worth the additional complexity.

Documenting

If a function implements Presto semantics, document it by adding an entry to one of the *.rst files in velox/docs/functions. Each file documents a set of related functions. E.g. math.rst contains all of the mathematical functions, while array.rst file contains all of the array functions. Within a file, functions are listed in alphabetical order.