lidavidm commented on code in PR #12590:
URL: https://github.com/apache/arrow/pull/12590#discussion_r860823509
##########
python/pyarrow/_compute.pyx:
##########
@@ -2275,3 +2279,205 @@ cdef CExpression _bind(Expression filter, Schema
schema) except *:
return GetResultValue(filter.unwrap().Bind(
deref(pyarrow_unwrap_schema(schema).get())))
+
+
+cdef class ScalarUdfContext:
+ """
+ Per-invocation function context/state.
+
+ This object will always be the first argument to a user-defined
+ function. It should not be used outside of a call to the function.
+ """
+
+ def __init__(self):
+ raise TypeError("Do not call {}'s constructor directly"
+ .format(self.__class__.__name__))
+
+ cdef void init(self, const CScalarUdfContext &c_context):
+ self.c_context = c_context
+
+ @property
+ def batch_length(self):
+ """
+ The common length of all input arguments (int).
+
+ In the case that all arguments are scalars, this value
+ is used to pass the "actual length" of the arguments,
+ e.g. because the scalar values are encoding a column
+ with a constant value.
+ """
+ return self.c_context.batch_length
+
+ @property
+ def memory_pool(self):
+ """
+ A memory pool for allocations (:class:`MemoryPool`).
+ """
+ return box_memory_pool(self.c_context.pool)
+
+
+cdef inline CFunctionDoc _make_function_doc(dict func_doc) except *:
+ """
+ Helper function to generate the FunctionDoc
+ This function accepts a dictionary and expect the
+ summary(str), description(str) and arg_names(List[str]) keys.
+ """
+ cdef:
+ CFunctionDoc f_doc
+ vector[c_string] c_arg_names
+
+ f_doc.summary = tobytes(func_doc["summary"])
+ f_doc.description = tobytes(func_doc["description"])
+ for arg_name in func_doc["arg_names"]:
+ c_arg_names.push_back(tobytes(arg_name))
+ f_doc.arg_names = c_arg_names
+ # UDFOptions integration:
+ # TODO: https://issues.apache.org/jira/browse/ARROW-16041
+ f_doc.options_class = b""
+ f_doc.options_required = False
+ return f_doc
+
+
+cdef object box_scalar_udf_context(const CScalarUdfContext& c_context):
+ cdef ScalarUdfContext context = ScalarUdfContext.__new__(ScalarUdfContext)
+ context.init(c_context)
+ return context
+
+
+cdef _scalar_udf_callback(user_function, const CScalarUdfContext& c_context,
inputs):
+ """
+ Helper callback function used to wrap the ScalarUdfContext from Python to
C++
+ execution.
+ """
+ context = box_scalar_udf_context(c_context)
+ return user_function(context, *inputs)
+
+
+def _get_scalar_udf_context(memory_pool, batch_length):
+ cdef CScalarUdfContext c_context
+ c_context.pool = maybe_unbox_memory_pool(memory_pool)
+ c_context.batch_length = batch_length
+ context = box_scalar_udf_context(c_context)
+ return context
+
+
+def register_scalar_function(func, function_name, function_doc, in_types,
+ out_type):
+ """
+ Register a user-defined scalar function.
+
+ A scalar function is a function that executes elementwise
+ operations on arrays or scalars. Also, a scalar function must
+ be computed row-by-row with no state where each output-row
+ is computed by only from it's corresponding input-row.
+ In other words, all argument arrays have the same length,
+ and the output array is of the same length as the arguments.
+ Scalar functions are the only functions allowed in query engine
+ expressions.
+
+ Parameters
+ ----------
+ func : callable
+ A callable implementing the user-defined function.
+ It must take arguments equal to the number of
+ in_types defined. It must return an Array or Scalar
+ matching the out_type. It must return a Scalar if
+ all arguments are scalar, else it must return an Array.
+
+ To define a varargs function, pass a callable that takes
+ varargs. The last in_type will be the type of all varargs
+ arguments.
+ function_name : str
+ Name of the function. This name must be globally unique.
+ function_doc : dict
+ A dictionary object with keys "summary" (str),
+ and "description" (str).
+ in_types : Dict[str, DataType]
+ A dictionarym mapping function argument names to
Review Comment:
```suggestion
A dictionary mapping function argument names to
```
##########
python/pyarrow/tests/test_udf.py:
##########
@@ -0,0 +1,496 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+
+import pytest
+
+import pyarrow as pa
+from pyarrow import compute as pc
+
+# UDFs are all tested with a dataset scan
+pytestmark = pytest.mark.dataset
+
+
+try:
+ import pyarrow.dataset as ds
+except ImportError:
+ ds = None
+
+global_batch_length = 10
+
+
+def mock_udf_context(batch_length=global_batch_length):
+ from pyarrow._compute import _get_scalar_udf_context
+ return _get_scalar_udf_context(pa.default_memory_pool(), batch_length)
+
+
[email protected](scope="session")
+def unary_func_fixture():
+ def unary_function(ctx, scalar1):
+ return pc.call_function("add", [scalar1, 1])
+ func_name = "y=x+k"
+ unary_doc = {"summary": "add function",
+ "description": "test add function"}
+ pc.register_scalar_function(unary_function,
+ func_name,
+ unary_doc,
+ {"array": pa.int64()},
+ pa.int64())
+ return unary_function, func_name
+
+
[email protected](scope="session")
+def binary_func_fixture():
+ def binary_function(ctx, m, x):
+ return pc.call_function("multiply", [m, x])
+ func_name = "y=mx"
+ binary_doc = {"summary": "y=mx",
+ "description": "find y from y = mx"}
+ pc.register_scalar_function(binary_function,
+ func_name,
+ binary_doc,
+ {"m": pa.int64(),
+ "x": pa.int64(),
+ },
+ pa.int64())
+ return binary_function, func_name
+
+
[email protected](scope="session")
+def ternary_func_fixture():
+ def ternary_function(ctx, m, x, c):
+ mx = pc.call_function("multiply", [m, x])
+ return pc.call_function("add", [mx, c])
+ ternary_doc = {"summary": "y=mx+c",
+ "description": "find y from y = mx + c"}
+ func_name = "y=mx+c"
+ pc.register_scalar_function(ternary_function,
+ func_name,
+ ternary_doc,
+ {
+ "array1": pa.int64(),
+ "array2": pa.int64(),
+ "array3": pa.int64(),
+ },
+ pa.int64())
+ return ternary_function, func_name
+
+
[email protected](scope="session")
+def varargs_func_fixture():
+ def varargs_function(ctx, *values):
+ base_val = values[:2]
+ res = pc.call_function("add", base_val)
+ for other_val in values[2:]:
+ res = pc.call_function("add", [res, other_val])
+ return res
+ func_name = "z=ax+by+c"
+ varargs_doc = {"summary": "z=ax+by+c",
+ "description": "find z from z = ax + by + c"
+ }
+ pc.register_scalar_function(varargs_function,
+ func_name,
+ varargs_doc,
+ {
+ "array1": pa.int64(),
+ "array2": pa.int64(),
+ "array3": pa.int64(),
+ "array4": pa.int64(),
+ "array5": pa.int64(),
+ },
+ pa.int64())
+ return varargs_function, func_name
+
+
[email protected](scope="session")
+def random_with_udf_ctx_func_fixture():
+ def random_with_udf_ctx(context, one, two):
+ return pc.add(one, two, memory_pool=context.memory_pool)
+
+ in_types = {"one": pa.int64(),
+ "two": pa.int64(),
+ }
+ func_doc = {
+ "summary": "test udf context",
+ "description": "udf context test"
+ }
+ func_name = "test_udf_context"
+ pc.register_scalar_function(random_with_udf_ctx,
+ func_name, func_doc,
+ in_types,
+ pa.int64())
+ return random_with_udf_ctx, func_name
+
+
[email protected](scope="session")
+def output_check_func_fixture():
+ # The objective of this fixture is to evaluate,
+ # how the UDF interface respond to unexpected
+ # output types. The types chosen at the test
+ # end are either of different Arrow data type
+ # or non-Arrow type.
+ def output_check(ctx, array):
+ ar = pc.call_function("add", [array, 1])
+ ar = ar.cast(pa.int32())
+ return ar
+ func_name = "test_output_value"
+ in_types = {"array": pa.int64()}
+ out_type = pa.int64()
+ doc = {
+ "summary": "add function scalar",
+ "description": "add function"
+ }
+ pc.register_scalar_function(output_check, func_name, doc,
+ in_types, out_type)
+ return output_check, func_name
+
+
[email protected](scope="session")
+def nullary_check_func_fixture():
+ # this needs to return array values
+ def nullary_check(context):
+ return pa.array([42] * context.batch_length, type=pa.int64(),
+ memory_pool=context.memory_pool)
+
+ func_doc = {
+ "summary": "random function",
+ "description": "generates a random value"
+ }
+ func_name = "test_random_func"
+ pc.register_scalar_function(nullary_check,
+ func_name,
+ func_doc,
+ {},
+ pa.int64())
+
+ return nullary_check, func_name
+
+
[email protected](scope="session")
+def output_python_type_func_fixture():
+ # This fixture helps to check the response
+ # when the function return value is not an Arrow
+ # defined data type. Instead here the returned value
+ # is of type int in Python.
+ def const_return(ctx, scalar):
+ return 42
+
+ func_name = "test_output_type"
+ in_types = {"array": pa.int64()}
+ out_type = pa.int64()
+ doc = {
+ "summary": "add function scalar",
+ "description": "add function"
+ }
+ pc.register_scalar_function(const_return, func_name, doc,
+ in_types, out_type)
+ return const_return, func_name
+
+
[email protected](scope="session")
+def varargs_check_func_fixture():
+ def varargs_check(ctx, *values):
+ base_val = values[:2]
+ res = pc.call_function("add", base_val)
+ for other_val in values[2:]:
+ res = pc.call_function("add", [res, other_val])
+ return res
+ func_name = "test_varargs_function"
+ in_types = {"array1": pa.int64(),
+ "array2": pa.int64(),
+ }
+ doc = {"summary": "n add function",
+ "description": "add N number of arrays"
+ }
+ pc.register_scalar_function(varargs_check, func_name, doc,
+ in_types, pa.int64())
+
+ return varargs_check, func_name
+
+
[email protected](scope="session")
+def raise_func_fixture():
+ def raise_func(ctx):
+ raise ValueError("Test function with raise")
+ func_name = "test_raise"
+ doc = {
+ "summary": "test function with raise",
+ "description": "function with a raise"
+ }
+ pc.register_scalar_function(raise_func, func_name, doc,
+ {}, pa.int64())
+ return raise_func, func_name
+
+
+def check_scalar_function(func_fixture,
+ input,
+ run_in_dataset=True,
+ batch_length=global_batch_length):
+ function, name = func_fixture
+ expected_output = function(mock_udf_context(batch_length), *input)
+ func = pc.get_function(name)
+ assert func.name == name
+
+ result = pc.call_function(name, input)
+
+ assert result == expected_output
+ if run_in_dataset:
+ field_names = [f'field{index}' for index, in_arr in input]
+ table = pa.Table.from_arrays(input, field_names)
+ dataset = ds.dataset(table)
+ func_args = [ds.field(field_name) for field_name in field_names]
+ result_table = dataset.to_table(
+ columns={'result': ds.field('')._call(name, func_args)})
+ assert result_table.column(0).chunks[0] == expected_output
+
+
+def test_scalar_udf_array_unary(unary_func_fixture):
+ check_scalar_function(unary_func_fixture,
+ [
+ pa.array([10, 20], pa.int64())
+ ]
+ )
+
+
+def test_scalar_udf_array_binary(binary_func_fixture):
+ check_scalar_function(binary_func_fixture,
+ [
+ pa.array([10, 20], pa.int64()),
+ pa.array([2, 4], pa.int64())
+ ]
+ )
+
+
+def test_scalar_udf_array_ternary(ternary_func_fixture):
+ check_scalar_function(ternary_func_fixture,
+ [
+ pa.array([10, 20], pa.int64()),
+ pa.array([2, 4], pa.int64()),
+ pa.array([5, 10], pa.int64())
+ ]
+ )
+
+
+def test_scalar_udf_array_varargs(varargs_func_fixture):
+ check_scalar_function(varargs_func_fixture,
+ [
+ pa.array([2, 3], pa.int64()),
+ pa.array([10, 20], pa.int64()),
+ pa.array([3, 7], pa.int64()),
+ pa.array([20, 30], pa.int64()),
+ pa.array([5, 10], pa.int64())
+ ],
+ mock_udf_context
Review Comment:
This shouldn't be here right?
##########
python/pyarrow/_compute.pyx:
##########
@@ -2275,3 +2279,205 @@ cdef CExpression _bind(Expression filter, Schema
schema) except *:
return GetResultValue(filter.unwrap().Bind(
deref(pyarrow_unwrap_schema(schema).get())))
+
+
+cdef class ScalarUdfContext:
+ """
+ Per-invocation function context/state.
+
+ This object will always be the first argument to a user-defined
+ function. It should not be used outside of a call to the function.
+ """
+
+ def __init__(self):
+ raise TypeError("Do not call {}'s constructor directly"
+ .format(self.__class__.__name__))
+
+ cdef void init(self, const CScalarUdfContext &c_context):
+ self.c_context = c_context
+
+ @property
+ def batch_length(self):
+ """
+ The common length of all input arguments (int).
+
+ In the case that all arguments are scalars, this value
+ is used to pass the "actual length" of the arguments,
+ e.g. because the scalar values are encoding a column
+ with a constant value.
+ """
+ return self.c_context.batch_length
+
+ @property
+ def memory_pool(self):
+ """
+ A memory pool for allocations (:class:`MemoryPool`).
+ """
+ return box_memory_pool(self.c_context.pool)
+
+
+cdef inline CFunctionDoc _make_function_doc(dict func_doc) except *:
+ """
+ Helper function to generate the FunctionDoc
+ This function accepts a dictionary and expect the
+ summary(str), description(str) and arg_names(List[str]) keys.
+ """
+ cdef:
+ CFunctionDoc f_doc
+ vector[c_string] c_arg_names
+
+ f_doc.summary = tobytes(func_doc["summary"])
+ f_doc.description = tobytes(func_doc["description"])
+ for arg_name in func_doc["arg_names"]:
+ c_arg_names.push_back(tobytes(arg_name))
+ f_doc.arg_names = c_arg_names
+ # UDFOptions integration:
+ # TODO: https://issues.apache.org/jira/browse/ARROW-16041
+ f_doc.options_class = b""
+ f_doc.options_required = False
+ return f_doc
+
+
+cdef object box_scalar_udf_context(const CScalarUdfContext& c_context):
+ cdef ScalarUdfContext context = ScalarUdfContext.__new__(ScalarUdfContext)
+ context.init(c_context)
+ return context
+
+
+cdef _scalar_udf_callback(user_function, const CScalarUdfContext& c_context,
inputs):
+ """
+ Helper callback function used to wrap the ScalarUdfContext from Python to
C++
+ execution.
+ """
+ context = box_scalar_udf_context(c_context)
+ return user_function(context, *inputs)
+
+
+def _get_scalar_udf_context(memory_pool, batch_length):
+ cdef CScalarUdfContext c_context
+ c_context.pool = maybe_unbox_memory_pool(memory_pool)
+ c_context.batch_length = batch_length
+ context = box_scalar_udf_context(c_context)
+ return context
+
+
+def register_scalar_function(func, function_name, function_doc, in_types,
+ out_type):
+ """
+ Register a user-defined scalar function.
+
+ A scalar function is a function that executes elementwise
+ operations on arrays or scalars. Also, a scalar function must
+ be computed row-by-row with no state where each output-row
+ is computed by only from it's corresponding input-row.
+ In other words, all argument arrays have the same length,
+ and the output array is of the same length as the arguments.
+ Scalar functions are the only functions allowed in query engine
+ expressions.
Review Comment:
```suggestion
A scalar function is a function that executes elementwise
operations on arrays or scalars, i.e. a scalar function must
be computed row-by-row with no state where each output row
is computed by only from its corresponding input row.
In other words, all argument arrays have the same length,
and the output array is of the same length as the arguments.
Scalar functions are the only functions allowed in query engine
expressions.
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
