cudf-polars-cu12 25.2.1__py3-none-any.whl → 25.4.0__py3-none-any.whl

cudf_polars/VERSION

@@ -1 +1 @@
-25.02.01
+25.04.00

cudf_polars/callback.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 
 """Callback for the polars collect function to execute on device."""
@@ -7,9 +7,10 @@ from __future__ import annotations
 
 import contextlib
 import os
+import time
 import warnings
 from functools import cache, partial
-from typing import TYPE_CHECKING, Literal
+from typing import TYPE_CHECKING, Literal, overload
 
 import nvtx
 
@@ -20,6 +21,8 @@ import rmm
 from rmm._cuda import gpu
 
 from cudf_polars.dsl.translate import Translator
+from cudf_polars.utils.timer import Timer
+from cudf_polars.utils.versions import POLARS_VERSION_LT_125
 
 if TYPE_CHECKING:
     from collections.abc import Generator
@@ -173,19 +176,53 @@ def set_device(device: int | None) -> Generator[int, None, None]:
         gpu.setDevice(previous)
 
 
+@overload
 def _callback(
     ir: IR,
     with_columns: list[str] | None,
     pyarrow_predicate: str | None,
     n_rows: int | None,
+    should_time: Literal[False],
     *,
     device: int | None,
     memory_resource: int | None,
     executor: Literal["pylibcudf", "dask-experimental"] | None,
-) -> pl.DataFrame:
+    timer: Timer | None,
+) -> pl.DataFrame: ...
+
+
+@overload
+def _callback(
+    ir: IR,
+    with_columns: list[str] | None,
+    pyarrow_predicate: str | None,
+    n_rows: int | None,
+    should_time: Literal[True],
+    *,
+    device: int | None,
+    memory_resource: int | None,
+    executor: Literal["pylibcudf", "dask-experimental"] | None,
+    timer: Timer | None,
+) -> tuple[pl.DataFrame, list[tuple[int, int, str]]]: ...
+
+
+def _callback(
+    ir: IR,
+    with_columns: list[str] | None,
+    pyarrow_predicate: str | None,
+    n_rows: int | None,
+    should_time: bool,  # noqa: FBT001
+    *,
+    device: int | None,
+    memory_resource: int | None,
+    executor: Literal["pylibcudf", "dask-experimental"] | None,
+    timer: Timer | None,
+):
     assert with_columns is None
     assert pyarrow_predicate is None
     assert n_rows is None
+    if timer is not None:
+        assert should_time
     with (
         nvtx.annotate(message="ExecuteIR", domain="cudf_polars"),
         # Device must be set before memory resource is obtained.
@@ -193,7 +230,11 @@ def _callback(
         set_memory_resource(memory_resource),
     ):
         if executor is None or executor == "pylibcudf":
-            return ir.evaluate(cache={}).to_polars()
+            df = ir.evaluate(cache={}, timer=timer).to_polars()
+            if timer is None:
+                return df
+            else:
+                return df, timer.timings
         elif executor == "dask-experimental":
             from cudf_polars.experimental.parallel import evaluate_dask
 
@@ -202,45 +243,9 @@ def _callback(
             raise ValueError(f"Unknown executor '{executor}'")
 
 
-def validate_config_options(config: dict) -> None:
-    """
-    Validate the configuration options for the GPU engine.
-
-    Parameters
-    ----------
-    config
-        Configuration options to validate.
-
-    Raises
-    ------
-    ValueError
-        If the configuration contains unsupported options.
-    """
-    if unsupported := (
-        config.keys()
-        - {"raise_on_fail", "parquet_options", "executor", "executor_options"}
-    ):
-        raise ValueError(
-            f"Engine configuration contains unsupported settings: {unsupported}"
-        )
-    assert {"chunked", "chunk_read_limit", "pass_read_limit"}.issuperset(
-        config.get("parquet_options", {})
-    )
-
-    # Validate executor_options
-    executor = config.get("executor", "pylibcudf")
-    if executor == "dask-experimental":
-        unsupported = config.get("executor_options", {}).keys() - {
-            "max_rows_per_partition",
-            "parquet_blocksize",
-        }
-    else:
-        unsupported = config.get("executor_options", {}).keys()
-    if unsupported:
-        raise ValueError(f"Unsupported executor_options for {executor}: {unsupported}")
-
-
-def execute_with_cudf(nt: NodeTraverser, *, config: GPUEngine) -> None:
+def execute_with_cudf(
+    nt: NodeTraverser, duration_since_start: int | None, *, config: GPUEngine
+) -> None:
     """
     A post optimization callback that attempts to execute the plan with cudf.
 
@@ -249,6 +254,10 @@ def execute_with_cudf(nt: NodeTraverser, *, config: GPUEngine) -> None:
     nt
         NodeTraverser
 
+    duration_since_start
+        Time since the user started executing the query (or None if no
+        profiling should occur).
+
     config
         GPUEngine configuration object
 
@@ -263,16 +272,21 @@ def execute_with_cudf(nt: NodeTraverser, *, config: GPUEngine) -> None:
     -----
     The NodeTraverser is mutated if the libcudf executor can handle the plan.
     """
+    if duration_since_start is None:
+        timer = None
+    else:
+        start = time.monotonic_ns()
+        timer = Timer(start - duration_since_start)
     device = config.device
     memory_resource = config.memory_resource
     raise_on_fail = config.config.get("raise_on_fail", False)
     executor = config.config.get("executor", None)
-    validate_config_options(config.config)
-
     with nvtx.annotate(message="ConvertIR", domain="cudf_polars"):
         translator = Translator(nt, config)
         ir = translator.translate_ir()
         ir_translation_errors = translator.errors
+        if timer is not None:
+            timer.store(start, time.monotonic_ns(), "gpu-ir-translation")
         if len(ir_translation_errors):
             # TODO: Display these errors in user-friendly way.
             # tracked in https://github.com/rapidsai/cudf/issues/17051
@@ -290,12 +304,30 @@ def execute_with_cudf(nt: NodeTraverser, *, config: GPUEngine) -> None:
             if raise_on_fail:
                 raise exception
         else:
-            nt.set_udf(
-                partial(
-                    _callback,
-                    ir,
-                    device=device,
-                    memory_resource=memory_resource,
-                    executor=executor,
+            if POLARS_VERSION_LT_125:  # pragma: no cover
+                nt.set_udf(
+                    partial(
+                        _callback,
+                        ir,
+                        should_time=False,
+                        device=device,
+                        memory_resource=memory_resource,
+                        executor=executor,
+                        timer=None,
+                    )
                 )
-            )
+            else:
+                nt.set_udf(
+                    partial(
+                        _callback,
+                        ir,
+                        device=device,
+                        memory_resource=memory_resource,
+                        executor=executor,
+                        timer=timer,
+                    )
+                )
+
+
+if POLARS_VERSION_LT_125:  # pragma: no cover
+    execute_with_cudf = partial(execute_with_cudf, duration_since_start=None)

cudf_polars/containers/column.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 
 """A column, with some properties."""
@@ -19,6 +19,7 @@ from pylibcudf.strings.convert.convert_integers import (
 )
 from pylibcudf.traits import is_floating_point
 
+from cudf_polars.utils import conversion
 from cudf_polars.utils.dtypes import is_order_preserving_cast
 
 if TYPE_CHECKING:
@@ -26,6 +27,8 @@ if TYPE_CHECKING:
 
     import polars as pl
 
+    from cudf_polars.typing import ColumnHeader, ColumnOptions, Slice
+
 __all__: list[str] = ["Column"]
 
 
@@ -51,10 +54,69 @@ class Column:
         name: str | None = None,
     ):
         self.obj = column
-        self.is_scalar = self.obj.size() == 1
+        self.is_scalar = self.size == 1
         self.name = name
         self.set_sorted(is_sorted=is_sorted, order=order, null_order=null_order)
 
+    @classmethod
+    def deserialize(
+        cls, header: ColumnHeader, frames: tuple[memoryview, plc.gpumemoryview]
+    ) -> Self:
+        """
+        Create a Column from a serialized representation returned by `.serialize()`.
+
+        Parameters
+        ----------
+        header
+            The (unpickled) metadata required to reconstruct the object.
+        frames
+            Two-tuple of frames (a memoryview and a gpumemoryview).
+
+        Returns
+        -------
+        Column
+            The deserialized Column.
+        """
+        packed_metadata, packed_gpu_data = frames
+        (plc_column,) = plc.contiguous_split.unpack_from_memoryviews(
+            packed_metadata, packed_gpu_data
+        ).columns()
+        return cls(plc_column, **header["column_kwargs"])
+
+    def serialize(
+        self,
+    ) -> tuple[ColumnHeader, tuple[memoryview, plc.gpumemoryview]]:
+        """
+        Serialize the Column into header and frames.
+
+        Follows the Dask serialization scheme with a picklable header (dict) and
+        a tuple of frames (in this case a contiguous host and device buffer).
+
+        To enable dask support, dask serializers must be registered
+
+            >>> from cudf_polars.experimental.dask_serialize import register
+            >>> register()
+
+        Returns
+        -------
+        header
+            A dict containing any picklable metadata required to reconstruct the object.
+        frames
+            Two-tuple of frames suitable for passing to `plc.contiguous_split.unpack_from_memoryviews`
+        """
+        packed = plc.contiguous_split.pack(plc.Table([self.obj]))
+        column_kwargs: ColumnOptions = {
+            "is_sorted": self.is_sorted,
+            "order": self.order,
+            "null_order": self.null_order,
+            "name": self.name,
+        }
+        header: ColumnHeader = {
+            "column_kwargs": column_kwargs,
+            "frame_count": 2,
+        }
+        return header, packed.release()
+
     @functools.cached_property
     def obj_scalar(self) -> plc.Scalar:
         """
@@ -70,9 +132,7 @@ class Column:
             If the column is not length-1.
         """
         if not self.is_scalar:
-            raise ValueError(
-                f"Cannot convert a column of length {self.obj.size()} to scalar"
-            )
+            raise ValueError(f"Cannot convert a column of length {self.size} to scalar")
         return plc.copying.get_element(self.obj, 0)
 
     def rename(self, name: str | None, /) -> Self:
@@ -242,7 +302,7 @@ class Column:
         -------
         Self with metadata set.
         """
-        if self.obj.size() <= 1:
+        if self.size <= 1:
             is_sorted = plc.types.Sorted.YES
         self.is_sorted = is_sorted
         self.order = order
@@ -268,7 +328,7 @@ class Column:
     def mask_nans(self) -> Self:
         """Return a shallow copy of self with nans masked out."""
         if plc.traits.is_floating_point(self.obj.type()):
-            old_count = self.obj.null_count()
+            old_count = self.null_count
             mask, new_count = plc.transform.nans_to_nulls(self.obj)
             result = type(self)(self.obj.with_mask(mask, new_count))
             if old_count == new_count:
@@ -288,3 +348,36 @@ class Column:
                 )
             ).as_py()
         return 0
+
+    @property
+    def size(self) -> int:
+        """Return the size of the column."""
+        return self.obj.size()
+
+    @property
+    def null_count(self) -> int:
+        """Return the number of Null values in the column."""
+        return self.obj.null_count()
+
+    def slice(self, zlice: Slice | None) -> Self:
+        """
+        Slice a column.
+
+        Parameters
+        ----------
+        zlice
+            optional, tuple of start and length, negative values of start
+            treated as for python indexing. If not provided, returns self.
+
+        Returns
+        -------
+        New column (if zlice is not None) otherwise self (if it is)
+        """
+        if zlice is None:
+            return self
+        (table,) = plc.copying.slice(
+            plc.Table([self.obj]),
+            conversion.from_polars_slice(zlice, num_rows=self.size),
+        )
+        (column,) = table.columns()
+        return type(self)(column, name=self.name).sorted_like(self)

cudf_polars/containers/dataframe.py

@@ -1,13 +1,12 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 
 """A dataframe, with some properties."""
 
 from __future__ import annotations
 
-import pickle
 from functools import cached_property
-from typing import TYPE_CHECKING, Any, cast
+from typing import TYPE_CHECKING, cast
 
 import pyarrow as pa
 
@@ -16,13 +15,15 @@ import polars as pl
 import pylibcudf as plc
 
 from cudf_polars.containers import Column
-from cudf_polars.utils import dtypes
+from cudf_polars.utils import conversion, dtypes
 
 if TYPE_CHECKING:
     from collections.abc import Iterable, Mapping, Sequence, Set
 
     from typing_extensions import Self
 
+    from cudf_polars.typing import ColumnOptions, DataFrameHeader, Slice
+
 
 __all__: list[str] = ["DataFrame"]
 
@@ -150,7 +151,7 @@ class DataFrame:
 
     @classmethod
     def deserialize(
-        cls, header: Mapping[str, Any], frames: tuple[memoryview, plc.gpumemoryview]
+        cls, header: DataFrameHeader, frames: tuple[memoryview, plc.gpumemoryview]
     ) -> Self:
         """
         Create a DataFrame from a serialized representation returned by `.serialize()`.
@@ -178,7 +179,7 @@ class DataFrame:
 
     def serialize(
         self,
-    ) -> tuple[Mapping[str, Any], tuple[memoryview, plc.gpumemoryview]]:
+    ) -> tuple[DataFrameHeader, tuple[memoryview, plc.gpumemoryview]]:
         """
         Serialize the table into header and frames.
 
@@ -187,20 +188,20 @@ class DataFrame:
 
         To enable dask support, dask serializers must be registered
 
-        >>> from cudf_polars.experimental.dask_serialize import register
-        >>> register()
+            >>> from cudf_polars.experimental.dask_serialize import register
+            >>> register()
 
         Returns
         -------
         header
             A dict containing any picklable metadata required to reconstruct the object.
         frames
-            Two-tuple of frames suitable for passing to `unpack_from_memoryviews`
+            Two-tuple of frames suitable for passing to `plc.contiguous_split.unpack_from_memoryviews`
         """
         packed = plc.contiguous_split.pack(self.table)
 
         # Keyword arguments for `Column.__init__`.
-        columns_kwargs = [
+        columns_kwargs: list[ColumnOptions] = [
             {
                 "is_sorted": col.is_sorted,
                 "order": col.order,
@@ -209,10 +210,8 @@ class DataFrame:
             }
             for col in self.columns
         ]
-        header = {
+        header: DataFrameHeader = {
             "columns_kwargs": columns_kwargs,
-            # Dask Distributed uses "type-serialized" to dispatch deserialization
-            "type-serialized": pickle.dumps(type(self)),
             "frame_count": 2,
         }
         return header, packed.release()
@@ -296,7 +295,7 @@ class DataFrame:
         table = plc.stream_compaction.apply_boolean_mask(self.table, mask.obj)
         return type(self).from_table(table, self.column_names).sorted_like(self)
 
-    def slice(self, zlice: tuple[int, int] | None) -> Self:
+    def slice(self, zlice: Slice | None) -> Self:
         """
         Slice a dataframe.
 
@@ -312,14 +311,7 @@ class DataFrame:
         """
         if zlice is None:
             return self
-        start, length = zlice
-        if start < 0:
-            start += self.num_rows
-        # Polars implementation wraps negative start by num_rows, then
-        # adds length to start to get the end, then clamps both to
-        # [0, num_rows)
-        end = start + length
-        start = max(min(start, self.num_rows), 0)
-        end = max(min(end, self.num_rows), 0)
-        (table,) = plc.copying.slice(self.table, [start, end])
+        (table,) = plc.copying.slice(
+            self.table, conversion.from_polars_slice(zlice, num_rows=self.num_rows)
+        )
         return type(self).from_table(table, self.column_names).sorted_like(self)

cudf_polars/dsl/expr.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 # TODO: remove need for this
 # ruff: noqa: D101
@@ -30,6 +30,7 @@ from cudf_polars.dsl.expressions.datetime import TemporalFunction
 from cudf_polars.dsl.expressions.literal import Literal, LiteralColumn
 from cudf_polars.dsl.expressions.rolling import GroupedRollingWindow, RollingWindow
 from cudf_polars.dsl.expressions.selection import Filter, Gather
+from cudf_polars.dsl.expressions.slicing import Slice
 from cudf_polars.dsl.expressions.sorting import Sort, SortBy
 from cudf_polars.dsl.expressions.string import StringFunction
 from cudf_polars.dsl.expressions.ternary import Ternary
@@ -53,6 +54,7 @@ __all__ = [
     "LiteralColumn",
     "NamedExpr",
     "RollingWindow",
+    "Slice",
     "Sort",
     "SortBy",
     "StringFunction",

cudf_polars/dsl/expressions/aggregation.py

@@ -172,7 +172,7 @@ class Agg(Expr):
             plc.Column.from_scalar(
                 plc.interop.from_arrow(
                     pa.scalar(
-                        column.obj.size() - column.obj.null_count(),
+                        column.size - column.null_count,
                         type=plc.interop.to_arrow(self.dtype),
                     ),
                 ),
@@ -181,7 +181,7 @@ class Agg(Expr):
         )
 
     def _sum(self, column: Column) -> Column:
-        if column.obj.size() == 0:
+        if column.size == 0 or column.null_count == column.size:
             return Column(
                 plc.Column.from_scalar(
                     plc.interop.from_arrow(
@@ -224,7 +224,7 @@ class Agg(Expr):
         return Column(plc.copying.slice(column.obj, [0, 1])[0])
 
     def _last(self, column: Column) -> Column:
-        n = column.obj.size()
+        n = column.size
         return Column(plc.copying.slice(column.obj, [n - 1, n])[0])
 
     def do_evaluate(

cudf_polars/dsl/expressions/binaryop.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 # TODO: remove need for this
 # ruff: noqa: D101
@@ -98,7 +98,7 @@ class BinOp(Expr):
         )
         lop = left.obj
         rop = right.obj
-        if left.obj.size() != right.obj.size():
+        if left.size != right.size:
             if left.is_scalar:
                 lop = left.obj_scalar
             elif right.is_scalar:

cudf_polars/dsl/expressions/boolean.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 # TODO: remove need for this
 # ruff: noqa: D101
@@ -191,7 +191,7 @@ class BooleanFunction(Expr):
             is_any = self.name is BooleanFunction.Name.Any
             agg = plc.aggregation.any() if is_any else plc.aggregation.all()
             result = plc.reduce.reduce(column.obj, agg, self.dtype)
-            if not ignore_nulls and column.obj.null_count() > 0:
+            if not ignore_nulls and column.null_count > 0:
                 #      Truth tables
                 #     Any         All
                 #   | F U T     | F U T
@@ -218,14 +218,14 @@ class BooleanFunction(Expr):
             (column,) = columns
             return Column(
                 plc.unary.is_nan(column.obj).with_mask(
-                    column.obj.null_mask(), column.obj.null_count()
+                    column.obj.null_mask(), column.null_count
                 )
             )
         elif self.name is BooleanFunction.Name.IsNotNan:
             (column,) = columns
             return Column(
                 plc.unary.is_not_nan(column.obj).with_mask(
-                    column.obj.null_mask(), column.obj.null_count()
+                    column.obj.null_mask(), column.null_count
                 )
             )
         elif self.name is BooleanFunction.Name.IsFirstDistinct:

cudf_polars/dsl/expressions/datetime.py

@@ -104,6 +104,14 @@ class TemporalFunction(Expr):
         Name.Nanosecond: plc.datetime.DatetimeComponent.NANOSECOND,
     }
 
+    _valid_ops: ClassVar[set[Name]] = {
+        *_COMPONENT_MAP.keys(),
+        Name.IsLeapYear,
+        Name.OrdinalDay,
+        Name.MonthStart,
+        Name.MonthEnd,
+    }
+
     def __init__(
         self,
         dtype: plc.DataType,
@@ -116,7 +124,7 @@ class TemporalFunction(Expr):
         self.name = name
         self.children = children
         self.is_pointwise = True
-        if self.name not in self._COMPONENT_MAP:
+        if self.name not in self._valid_ops:
             raise NotImplementedError(f"Temporal function {self.name}")
 
     def do_evaluate(
@@ -132,6 +140,36 @@ class TemporalFunction(Expr):
             for child in self.children
         ]
         (column,) = columns
+        if self.name is TemporalFunction.Name.MonthStart:
+            ends = plc.datetime.last_day_of_month(column.obj)
+            days_to_subtract = plc.datetime.days_in_month(column.obj)
+            # must subtract 1 to avoid rolling over to the previous month
+            days_to_subtract = plc.binaryop.binary_operation(
+                days_to_subtract,
+                plc.interop.from_arrow(pa.scalar(1, type=pa.int32())),
+                plc.binaryop.BinaryOperator.SUB,
+                plc.DataType(plc.TypeId.DURATION_DAYS),
+            )
+            result = plc.binaryop.binary_operation(
+                ends,
+                days_to_subtract,
+                plc.binaryop.BinaryOperator.SUB,
+                column.obj.type(),
+            )
+
+            return Column(result)
+        if self.name is TemporalFunction.Name.MonthEnd:
+            return Column(
+                plc.unary.cast(
+                    plc.datetime.last_day_of_month(column.obj), column.obj.type()
+                )
+            )
+        if self.name is TemporalFunction.Name.IsLeapYear:
+            return Column(
+                plc.datetime.is_leap_year(column.obj),
+            )
+        if self.name is TemporalFunction.Name.OrdinalDay:
+            return Column(plc.datetime.day_of_year(column.obj))
         if self.name is TemporalFunction.Name.Microsecond:
             millis = plc.datetime.extract_datetime_component(
                 column.obj, plc.datetime.DatetimeComponent.MILLISECOND

cudf_polars/dsl/expressions/literal.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 # TODO: remove need for this
 # ruff: noqa: D101
@@ -8,21 +8,16 @@ from __future__ import annotations
 
 from typing import TYPE_CHECKING, Any
 
-import pyarrow as pa
-
 import pylibcudf as plc
 
 from cudf_polars.containers import Column
 from cudf_polars.dsl.expressions.base import AggInfo, ExecutionContext, Expr
-from cudf_polars.utils import dtypes
 
 if TYPE_CHECKING:
     from collections.abc import Hashable, Mapping
 
     import pyarrow as pa
 
-    import polars as pl
-
     from cudf_polars.containers import DataFrame
 
 __all__ = ["Literal", "LiteralColumn"]
@@ -61,10 +56,9 @@ class LiteralColumn(Expr):
     _non_child = ("dtype", "value")
     value: pa.Array[Any]
 
-    def __init__(self, dtype: plc.DataType, value: pl.Series) -> None:
+    def __init__(self, dtype: plc.DataType, value: pa.Array) -> None:
         self.dtype = dtype
-        data = value.to_arrow()
-        self.value = data.cast(dtypes.downcast_arrow_lists(data.type))
+        self.value = value
         self.children = ()
         self.is_pointwise = True