team-10/venv/Lib/site-packages/narwhals/_arrow/series.py
2025-08-02 02:00:33 +02:00

1183 lines
44 KiB
Python

from __future__ import annotations
from typing import TYPE_CHECKING, Any, Literal, cast, overload
import pyarrow as pa
import pyarrow.compute as pc
from narwhals._arrow.series_cat import ArrowSeriesCatNamespace
from narwhals._arrow.series_dt import ArrowSeriesDateTimeNamespace
from narwhals._arrow.series_list import ArrowSeriesListNamespace
from narwhals._arrow.series_str import ArrowSeriesStringNamespace
from narwhals._arrow.series_struct import ArrowSeriesStructNamespace
from narwhals._arrow.utils import (
cast_for_truediv,
chunked_array,
extract_native,
floordiv_compat,
lit,
narwhals_to_native_dtype,
native_to_narwhals_dtype,
nulls_like,
pad_series,
zeros,
)
from narwhals._compliant import EagerSeries, EagerSeriesHist
from narwhals._expression_parsing import ExprKind
from narwhals._typing_compat import assert_never
from narwhals._utils import (
Implementation,
generate_temporary_column_name,
is_list_of,
not_implemented,
)
from narwhals.dependencies import is_numpy_array_1d
from narwhals.exceptions import InvalidOperationError, ShapeError
if TYPE_CHECKING:
from collections.abc import Iterable, Iterator, Mapping, Sequence
from types import ModuleType
import pandas as pd
import polars as pl
from typing_extensions import Self, TypeAlias, TypeIs
from narwhals._arrow.dataframe import ArrowDataFrame
from narwhals._arrow.namespace import ArrowNamespace
from narwhals._arrow.typing import ( # type: ignore[attr-defined]
ArrayAny,
ArrayOrChunkedArray,
ArrayOrScalar,
ChunkedArrayAny,
Incomplete,
NullPlacement,
Order,
ScalarAny,
TieBreaker,
_AsPyType,
_BasicDataType,
)
from narwhals._compliant.series import HistData
from narwhals._utils import Version, _LimitedContext
from narwhals.dtypes import DType
from narwhals.typing import (
ClosedInterval,
FillNullStrategy,
Into1DArray,
IntoDType,
NonNestedLiteral,
NumericLiteral,
PythonLiteral,
RankMethod,
RollingInterpolationMethod,
SizedMultiIndexSelector,
TemporalLiteral,
_1DArray,
_2DArray,
_SliceIndex,
)
ArrowHistData: TypeAlias = (
"HistData[ChunkedArrayAny, list[ScalarAny] | pa.Int64Array | list[float]]"
)
# TODO @dangotbanned: move into `_arrow.utils`
# Lots of modules are importing inline
@overload
def maybe_extract_py_scalar(
value: pa.Scalar[_BasicDataType[_AsPyType]],
return_py_scalar: bool, # noqa: FBT001
) -> _AsPyType: ...
@overload
def maybe_extract_py_scalar(
value: pa.Scalar[pa.StructType],
return_py_scalar: bool, # noqa: FBT001
) -> list[dict[str, Any]]: ...
@overload
def maybe_extract_py_scalar(
value: pa.Scalar[pa.ListType[_BasicDataType[_AsPyType]]],
return_py_scalar: bool, # noqa: FBT001
) -> list[_AsPyType]: ...
@overload
def maybe_extract_py_scalar(
value: pa.Scalar[Any] | Any,
return_py_scalar: bool, # noqa: FBT001
) -> Any: ...
def maybe_extract_py_scalar(value: Any, return_py_scalar: bool) -> Any: # noqa: FBT001
if TYPE_CHECKING:
return value.as_py()
if return_py_scalar:
return getattr(value, "as_py", lambda: value)()
return value
class ArrowSeries(EagerSeries["ChunkedArrayAny"]):
_implementation = Implementation.PYARROW
def __init__(
self, native_series: ChunkedArrayAny, *, name: str, version: Version
) -> None:
self._name = name
self._native_series: ChunkedArrayAny = native_series
self._version = version
self._broadcast = False
@property
def native(self) -> ChunkedArrayAny:
return self._native_series
def _with_version(self, version: Version) -> Self:
return self.__class__(self.native, name=self._name, version=version)
def _with_native(
self, series: ArrayOrScalar, *, preserve_broadcast: bool = False
) -> Self:
result = self.from_native(chunked_array(series), name=self.name, context=self)
if preserve_broadcast:
result._broadcast = self._broadcast
return result
@classmethod
def from_iterable(
cls,
data: Iterable[Any],
*,
context: _LimitedContext,
name: str = "",
dtype: IntoDType | None = None,
) -> Self:
version = context._version
dtype_pa = narwhals_to_native_dtype(dtype, version) if dtype else None
return cls.from_native(
chunked_array([data], dtype_pa), name=name, context=context
)
def _from_scalar(self, value: Any) -> Self:
if hasattr(value, "as_py"):
value = value.as_py()
return super()._from_scalar(value)
@staticmethod
def _is_native(obj: ChunkedArrayAny | Any) -> TypeIs[ChunkedArrayAny]:
return isinstance(obj, pa.ChunkedArray)
@classmethod
def from_native(
cls, data: ChunkedArrayAny, /, *, context: _LimitedContext, name: str = ""
) -> Self:
return cls(data, version=context._version, name=name)
@classmethod
def from_numpy(cls, data: Into1DArray, /, *, context: _LimitedContext) -> Self:
return cls.from_iterable(
data if is_numpy_array_1d(data) else [data], context=context
)
@classmethod
def _align_full_broadcast(cls, *series: Self) -> Sequence[Self]:
lengths = [len(s) for s in series]
max_length = max(lengths)
fast_path = all(_len == max_length for _len in lengths)
if fast_path:
return series
reshaped = []
for s in series:
if s._broadcast:
compliant = s._with_native(pa.repeat(s.native[0], max_length))
elif (actual_len := len(s)) != max_length:
msg = f"Expected object of length {max_length}, got {actual_len}."
raise ShapeError(msg)
else:
compliant = s
reshaped.append(compliant)
return reshaped
def __narwhals_namespace__(self) -> ArrowNamespace:
from narwhals._arrow.namespace import ArrowNamespace
return ArrowNamespace(version=self._version)
def __eq__(self, other: object) -> Self: # type: ignore[override]
other = cast("PythonLiteral | ArrowSeries | None", other)
ser, rhs = extract_native(self, other)
return self._with_native(pc.equal(ser, rhs))
def __ne__(self, other: object) -> Self: # type: ignore[override]
other = cast("PythonLiteral | ArrowSeries | None", other)
ser, rhs = extract_native(self, other)
return self._with_native(pc.not_equal(ser, rhs))
def __ge__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.greater_equal(ser, other))
def __gt__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.greater(ser, other))
def __le__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.less_equal(ser, other))
def __lt__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.less(ser, other))
def __and__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.and_kleene(ser, other)) # type: ignore[arg-type]
def __rand__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.and_kleene(other, ser)) # type: ignore[arg-type]
def __or__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.or_kleene(ser, other)) # type: ignore[arg-type]
def __ror__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.or_kleene(other, ser)) # type: ignore[arg-type]
def __add__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.add(ser, other))
def __radd__(self, other: Any) -> Self:
return self + other
def __sub__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.subtract(ser, other))
def __rsub__(self, other: Any) -> Self:
return (self - other) * (-1)
def __mul__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.multiply(ser, other))
def __rmul__(self, other: Any) -> Self:
return self * other
def __pow__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.power(ser, other))
def __rpow__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.power(other, ser))
def __floordiv__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(floordiv_compat(ser, other))
def __rfloordiv__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(floordiv_compat(other, ser))
def __truediv__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.divide(*cast_for_truediv(ser, other))) # type: ignore[type-var]
def __rtruediv__(self, other: Any) -> Self:
ser, other = extract_native(self, other)
return self._with_native(pc.divide(*cast_for_truediv(other, ser))) # type: ignore[type-var]
def __mod__(self, other: Any) -> Self:
floor_div = (self // other).native
ser, other = extract_native(self, other)
res = pc.subtract(ser, pc.multiply(floor_div, other))
return self._with_native(res)
def __rmod__(self, other: Any) -> Self:
floor_div = (other // self).native
ser, other = extract_native(self, other)
res = pc.subtract(other, pc.multiply(floor_div, ser))
return self._with_native(res)
def __invert__(self) -> Self:
return self._with_native(pc.invert(self.native))
@property
def _type(self) -> pa.DataType:
return self.native.type
def len(self, *, _return_py_scalar: bool = True) -> int:
return maybe_extract_py_scalar(len(self.native), _return_py_scalar)
def filter(self, predicate: ArrowSeries | list[bool | None]) -> Self:
other_native: Any
if not is_list_of(predicate, bool):
_, other_native = extract_native(self, predicate)
else:
other_native = predicate
return self._with_native(self.native.filter(other_native))
def mean(self, *, _return_py_scalar: bool = True) -> float:
return maybe_extract_py_scalar(pc.mean(self.native), _return_py_scalar)
def median(self, *, _return_py_scalar: bool = True) -> float:
if not self.dtype.is_numeric():
msg = "`median` operation not supported for non-numeric input type."
raise InvalidOperationError(msg)
return maybe_extract_py_scalar(
pc.approximate_median(self.native), _return_py_scalar
)
def min(self, *, _return_py_scalar: bool = True) -> Any:
return maybe_extract_py_scalar(pc.min(self.native), _return_py_scalar)
def max(self, *, _return_py_scalar: bool = True) -> Any:
return maybe_extract_py_scalar(pc.max(self.native), _return_py_scalar)
def arg_min(self, *, _return_py_scalar: bool = True) -> int:
index_min = pc.index(self.native, pc.min(self.native))
return maybe_extract_py_scalar(index_min, _return_py_scalar)
def arg_max(self, *, _return_py_scalar: bool = True) -> int:
index_max = pc.index(self.native, pc.max(self.native))
return maybe_extract_py_scalar(index_max, _return_py_scalar)
def sum(self, *, _return_py_scalar: bool = True) -> float:
return maybe_extract_py_scalar(
pc.sum(self.native, min_count=0), _return_py_scalar
)
def drop_nulls(self) -> Self:
return self._with_native(self.native.drop_null())
def shift(self, n: int) -> Self:
if n > 0:
arrays = [nulls_like(n, self), *self.native[:-n].chunks]
elif n < 0:
arrays = [*self.native[-n:].chunks, nulls_like(-n, self)]
else:
return self._with_native(self.native)
return self._with_native(pa.concat_arrays(arrays))
def std(self, ddof: int, *, _return_py_scalar: bool = True) -> float:
return maybe_extract_py_scalar(
pc.stddev(self.native, ddof=ddof), _return_py_scalar
)
def var(self, ddof: int, *, _return_py_scalar: bool = True) -> float:
return maybe_extract_py_scalar(
pc.variance(self.native, ddof=ddof), _return_py_scalar
)
def skew(self, *, _return_py_scalar: bool = True) -> float | None:
ser_not_null = self.native.drop_null()
if len(ser_not_null) == 0:
return None
elif len(ser_not_null) == 1:
return float("nan")
elif len(ser_not_null) == 2:
return 0.0
else:
m = pc.subtract(ser_not_null, pc.mean(ser_not_null))
m2 = pc.mean(pc.power(m, lit(2)))
m3 = pc.mean(pc.power(m, lit(3)))
biased_population_skewness = pc.divide(m3, pc.power(m2, lit(1.5)))
return maybe_extract_py_scalar(biased_population_skewness, _return_py_scalar)
def kurtosis(self, *, _return_py_scalar: bool = True) -> float | None:
ser_not_null = self.native.drop_null()
if len(ser_not_null) == 0:
return None
elif len(ser_not_null) == 1:
return float("nan")
else:
m = pc.subtract(ser_not_null, pc.mean(ser_not_null))
m2 = pc.mean(pc.power(m, lit(2)))
m4 = pc.mean(pc.power(m, lit(4)))
k = pc.subtract(pc.divide(m4, pc.power(m2, lit(2))), lit(3))
return maybe_extract_py_scalar(k, _return_py_scalar)
def count(self, *, _return_py_scalar: bool = True) -> int:
return maybe_extract_py_scalar(pc.count(self.native), _return_py_scalar)
def n_unique(self, *, _return_py_scalar: bool = True) -> int:
return maybe_extract_py_scalar(
pc.count(self.native.unique(), mode="all"), _return_py_scalar
)
def __native_namespace__(self) -> ModuleType:
if self._implementation is Implementation.PYARROW:
return self._implementation.to_native_namespace()
msg = f"Expected pyarrow, got: {type(self._implementation)}" # pragma: no cover
raise AssertionError(msg)
@property
def name(self) -> str:
return self._name
def _gather(self, rows: SizedMultiIndexSelector[ChunkedArrayAny]) -> Self:
if len(rows) == 0:
return self._with_native(self.native.slice(0, 0))
if self._backend_version < (18,) and isinstance(rows, tuple):
rows = list(rows)
return self._with_native(self.native.take(rows))
def _gather_slice(self, rows: _SliceIndex | range) -> Self:
start = rows.start or 0
stop = rows.stop if rows.stop is not None else len(self.native)
if start < 0:
start = len(self.native) + start
if stop < 0:
stop = len(self.native) + stop
if rows.step is not None and rows.step != 1:
msg = "Slicing with step is not supported on PyArrow tables"
raise NotImplementedError(msg)
return self._with_native(self.native.slice(start, stop - start))
def scatter(self, indices: int | Sequence[int], values: Any) -> Self:
import numpy as np # ignore-banned-import
values_native: ArrayAny
if isinstance(indices, int):
indices_native = pa.array([indices])
values_native = pa.array([values])
else:
# TODO(unassigned): we may also want to let `indices` be a Series.
# https://github.com/narwhals-dev/narwhals/issues/2155
indices_native = pa.array(indices)
if isinstance(values, self.__class__):
values_native = values.native.combine_chunks()
else:
# NOTE: Requires fixes in https://github.com/zen-xu/pyarrow-stubs/pull/209
pa_array: Incomplete = pa.array
values_native = pa_array(values)
sorting_indices = pc.sort_indices(indices_native)
indices_native = indices_native.take(sorting_indices)
values_native = values_native.take(sorting_indices)
mask: _1DArray = np.zeros(self.len(), dtype=bool)
mask[indices_native] = True
# NOTE: Multiple issues
# - Missing `values` type
# - `mask` accepts a `np.ndarray`, but not mentioned in stubs
# - Missing `replacements` type
# - Missing return type
pc_replace_with_mask: Incomplete = pc.replace_with_mask
return self._with_native(pc_replace_with_mask(self.native, mask, values_native))
def to_list(self) -> list[Any]:
return self.native.to_pylist()
def __array__(self, dtype: Any = None, *, copy: bool | None = None) -> _1DArray:
return self.native.__array__(dtype=dtype, copy=copy)
def to_numpy(self, dtype: Any = None, *, copy: bool | None = None) -> _1DArray:
return self.native.to_numpy()
def alias(self, name: str) -> Self:
result = self.__class__(self.native, name=name, version=self._version)
result._broadcast = self._broadcast
return result
@property
def dtype(self) -> DType:
return native_to_narwhals_dtype(self.native.type, self._version)
def abs(self) -> Self:
return self._with_native(pc.abs(self.native))
def cum_sum(self, *, reverse: bool) -> Self:
cum_sum = pc.cumulative_sum
result = (
cum_sum(self.native, skip_nulls=True)
if not reverse
else cum_sum(self.native[::-1], skip_nulls=True)[::-1]
)
return self._with_native(result)
def round(self, decimals: int) -> Self:
return self._with_native(
pc.round(self.native, decimals, round_mode="half_towards_infinity")
)
def diff(self) -> Self:
return self._with_native(pc.pairwise_diff(self.native.combine_chunks()))
def any(self, *, _return_py_scalar: bool = True) -> bool:
return maybe_extract_py_scalar(
pc.any(self.native, min_count=0), _return_py_scalar
)
def all(self, *, _return_py_scalar: bool = True) -> bool:
return maybe_extract_py_scalar(
pc.all(self.native, min_count=0), _return_py_scalar
)
def is_between(
self, lower_bound: Any, upper_bound: Any, closed: ClosedInterval
) -> Self:
_, lower_bound = extract_native(self, lower_bound)
_, upper_bound = extract_native(self, upper_bound)
if closed == "left":
ge = pc.greater_equal(self.native, lower_bound)
lt = pc.less(self.native, upper_bound)
res = pc.and_kleene(ge, lt)
elif closed == "right":
gt = pc.greater(self.native, lower_bound)
le = pc.less_equal(self.native, upper_bound)
res = pc.and_kleene(gt, le)
elif closed == "none":
gt = pc.greater(self.native, lower_bound)
lt = pc.less(self.native, upper_bound)
res = pc.and_kleene(gt, lt)
elif closed == "both":
ge = pc.greater_equal(self.native, lower_bound)
le = pc.less_equal(self.native, upper_bound)
res = pc.and_kleene(ge, le)
else:
assert_never(closed)
return self._with_native(res)
def is_null(self) -> Self:
return self._with_native(self.native.is_null(), preserve_broadcast=True)
def is_nan(self) -> Self:
return self._with_native(pc.is_nan(self.native), preserve_broadcast=True)
def cast(self, dtype: IntoDType) -> Self:
data_type = narwhals_to_native_dtype(dtype, self._version)
return self._with_native(pc.cast(self.native, data_type), preserve_broadcast=True)
def null_count(self, *, _return_py_scalar: bool = True) -> int:
return maybe_extract_py_scalar(self.native.null_count, _return_py_scalar)
def head(self, n: int) -> Self:
if n >= 0:
return self._with_native(self.native.slice(0, n))
else:
num_rows = len(self)
return self._with_native(self.native.slice(0, max(0, num_rows + n)))
def tail(self, n: int) -> Self:
if n >= 0:
num_rows = len(self)
return self._with_native(self.native.slice(max(0, num_rows - n)))
else:
return self._with_native(self.native.slice(abs(n)))
def is_in(self, other: Any) -> Self:
if self._is_native(other):
value_set: ArrayOrChunkedArray = other
else:
value_set = pa.array(other)
return self._with_native(pc.is_in(self.native, value_set=value_set))
def arg_true(self) -> Self:
import numpy as np # ignore-banned-import
res = np.flatnonzero(self.native)
return self.from_iterable(res, name=self.name, context=self)
def item(self, index: int | None = None) -> Any:
if index is None:
if len(self) != 1:
msg = (
"can only call '.item()' if the Series is of length 1,"
f" or an explicit index is provided (Series is of length {len(self)})"
)
raise ValueError(msg)
return maybe_extract_py_scalar(self.native[0], return_py_scalar=True)
return maybe_extract_py_scalar(self.native[index], return_py_scalar=True)
def value_counts(
self, *, sort: bool, parallel: bool, name: str | None, normalize: bool
) -> ArrowDataFrame:
"""Parallel is unused, exists for compatibility."""
from narwhals._arrow.dataframe import ArrowDataFrame
index_name_ = "index" if self._name is None else self._name
value_name_ = name or ("proportion" if normalize else "count")
val_counts = pc.value_counts(self.native)
values = val_counts.field("values")
counts = cast("ChunkedArrayAny", val_counts.field("counts"))
if normalize:
arrays = [values, pc.divide(*cast_for_truediv(counts, pc.sum(counts)))]
else:
arrays = [values, counts]
val_count = pa.Table.from_arrays(arrays, names=[index_name_, value_name_])
if sort:
val_count = val_count.sort_by([(value_name_, "descending")])
return ArrowDataFrame(
val_count, version=self._version, validate_column_names=True
)
def zip_with(self, mask: Self, other: Self) -> Self:
cond = mask.native.combine_chunks()
return self._with_native(pc.if_else(cond, self.native, other.native))
def sample(
self,
n: int | None,
*,
fraction: float | None,
with_replacement: bool,
seed: int | None,
) -> Self:
import numpy as np # ignore-banned-import
num_rows = len(self)
if n is None and fraction is not None:
n = int(num_rows * fraction)
rng = np.random.default_rng(seed=seed)
idx = np.arange(num_rows)
mask = rng.choice(idx, size=n, replace=with_replacement)
return self._with_native(self.native.take(mask))
def fill_null(
self,
value: Self | NonNestedLiteral,
strategy: FillNullStrategy | None,
limit: int | None,
) -> Self:
import numpy as np # ignore-banned-import
def fill_aux(
arr: ChunkedArrayAny, limit: int, direction: FillNullStrategy | None
) -> ArrayAny:
# this algorithm first finds the indices of the valid values to fill all the null value positions
# then it calculates the distance of each new index and the original index
# if the distance is equal to or less than the limit and the original value is null, it is replaced
valid_mask = pc.is_valid(arr)
indices = pa.array(np.arange(len(arr)), type=pa.int64())
if direction == "forward":
valid_index = np.maximum.accumulate(np.where(valid_mask, indices, -1))
distance = indices - valid_index
else:
valid_index = np.minimum.accumulate(
np.where(valid_mask[::-1], indices[::-1], len(arr))
)[::-1]
distance = valid_index - indices
return pc.if_else(
pc.and_(pc.is_null(arr), pc.less_equal(distance, lit(limit))), # pyright: ignore[reportArgumentType, reportCallIssue]
arr.take(valid_index),
arr,
)
if value is not None:
_, native_value = extract_native(self, value)
series: ArrayOrScalar = pc.fill_null(self.native, native_value)
elif limit is None:
fill_func = (
pc.fill_null_forward if strategy == "forward" else pc.fill_null_backward
)
series = fill_func(self.native)
else:
series = fill_aux(self.native, limit, strategy)
return self._with_native(series, preserve_broadcast=True)
def to_frame(self) -> ArrowDataFrame:
from narwhals._arrow.dataframe import ArrowDataFrame
df = pa.Table.from_arrays([self.native], names=[self.name])
return ArrowDataFrame(df, version=self._version, validate_column_names=False)
def to_pandas(self) -> pd.Series[Any]:
import pandas as pd # ignore-banned-import()
return pd.Series(self.native, name=self.name)
def to_polars(self) -> pl.Series:
import polars as pl # ignore-banned-import
return cast("pl.Series", pl.from_arrow(self.native))
def is_unique(self) -> ArrowSeries:
return self.to_frame().is_unique().alias(self.name)
def is_first_distinct(self) -> Self:
import numpy as np # ignore-banned-import
row_number = pa.array(np.arange(len(self)))
col_token = generate_temporary_column_name(n_bytes=8, columns=[self.name])
first_distinct_index = (
pa.Table.from_arrays([self.native], names=[self.name])
.append_column(col_token, row_number)
.group_by(self.name)
.aggregate([(col_token, "min")])
.column(f"{col_token}_min")
)
return self._with_native(pc.is_in(row_number, first_distinct_index))
def is_last_distinct(self) -> Self:
import numpy as np # ignore-banned-import
row_number = pa.array(np.arange(len(self)))
col_token = generate_temporary_column_name(n_bytes=8, columns=[self.name])
last_distinct_index = (
pa.Table.from_arrays([self.native], names=[self.name])
.append_column(col_token, row_number)
.group_by(self.name)
.aggregate([(col_token, "max")])
.column(f"{col_token}_max")
)
return self._with_native(pc.is_in(row_number, last_distinct_index))
def is_sorted(self, *, descending: bool) -> bool:
if not isinstance(descending, bool):
msg = f"argument 'descending' should be boolean, found {type(descending)}"
raise TypeError(msg)
if descending:
result = pc.all(pc.greater_equal(self.native[:-1], self.native[1:]))
else:
result = pc.all(pc.less_equal(self.native[:-1], self.native[1:]))
return maybe_extract_py_scalar(result, return_py_scalar=True)
def unique(self, *, maintain_order: bool) -> Self:
# TODO(marco): `pc.unique` seems to always maintain order, is that guaranteed?
return self._with_native(self.native.unique())
def replace_strict(
self,
old: Sequence[Any] | Mapping[Any, Any],
new: Sequence[Any],
*,
return_dtype: IntoDType | None,
) -> Self:
# https://stackoverflow.com/a/79111029/4451315
idxs = pc.index_in(self.native, pa.array(old))
result_native = pc.take(pa.array(new), idxs)
if return_dtype is not None:
result_native.cast(narwhals_to_native_dtype(return_dtype, self._version))
result = self._with_native(result_native)
if result.is_null().sum() != self.is_null().sum():
msg = (
"replace_strict did not replace all non-null values.\n\n"
"The following did not get replaced: "
f"{self.filter(~self.is_null() & result.is_null()).unique(maintain_order=False).to_list()}"
)
raise ValueError(msg)
return result
def sort(self, *, descending: bool, nulls_last: bool) -> Self:
order: Order = "descending" if descending else "ascending"
null_placement: NullPlacement = "at_end" if nulls_last else "at_start"
sorted_indices = pc.array_sort_indices(
self.native, order=order, null_placement=null_placement
)
return self._with_native(self.native.take(sorted_indices))
def to_dummies(self, *, separator: str, drop_first: bool) -> ArrowDataFrame:
import numpy as np # ignore-banned-import
from narwhals._arrow.dataframe import ArrowDataFrame
name = self._name
# NOTE: stub is missing attributes (https://arrow.apache.org/docs/python/generated/pyarrow.DictionaryArray.html)
da: Incomplete = self.native.combine_chunks().dictionary_encode("encode")
columns: _2DArray = np.zeros((len(da.dictionary), len(da)), np.int8)
columns[da.indices, np.arange(len(da))] = 1
null_col_pa, null_col_pl = f"{name}{separator}None", f"{name}{separator}null"
cols = [
{null_col_pa: null_col_pl}.get(
f"{name}{separator}{v}", f"{name}{separator}{v}"
)
for v in da.dictionary
]
output_order = (
[
null_col_pl,
*sorted([c for c in cols if c != null_col_pl])[int(drop_first) :],
]
if null_col_pl in cols
else sorted(cols)[int(drop_first) :]
)
return ArrowDataFrame(
pa.Table.from_arrays(columns, names=cols),
version=self._version,
validate_column_names=True,
).simple_select(*output_order)
def quantile(
self,
quantile: float,
interpolation: RollingInterpolationMethod,
*,
_return_py_scalar: bool = True,
) -> float:
return maybe_extract_py_scalar(
pc.quantile(self.native, q=quantile, interpolation=interpolation)[0],
_return_py_scalar,
)
def gather_every(self, n: int, offset: int = 0) -> Self:
return self._with_native(self.native[offset::n])
def clip(
self,
lower_bound: Self | NumericLiteral | TemporalLiteral | None,
upper_bound: Self | NumericLiteral | TemporalLiteral | None,
) -> Self:
_, lower = (
extract_native(self, lower_bound) if lower_bound is not None else (None, None)
)
_, upper = (
extract_native(self, upper_bound) if upper_bound is not None else (None, None)
)
if lower is None:
return self._with_native(pc.min_element_wise(self.native, upper))
if upper is None:
return self._with_native(pc.max_element_wise(self.native, lower))
return self._with_native(
pc.max_element_wise(pc.min_element_wise(self.native, upper), lower)
)
def to_arrow(self) -> ArrayAny:
return self.native.combine_chunks()
def mode(self) -> ArrowSeries:
plx = self.__narwhals_namespace__()
col_token = generate_temporary_column_name(n_bytes=8, columns=[self.name])
counts = self.value_counts(
name=col_token, normalize=False, sort=False, parallel=False
)
return counts.filter(
plx.col(col_token)
== plx.col(col_token).max().broadcast(kind=ExprKind.AGGREGATION)
).get_column(self.name)
def is_finite(self) -> Self:
return self._with_native(pc.is_finite(self.native))
def cum_count(self, *, reverse: bool) -> Self:
dtypes = self._version.dtypes
return (~self.is_null()).cast(dtypes.UInt32()).cum_sum(reverse=reverse)
def cum_min(self, *, reverse: bool) -> Self:
result = (
pc.cumulative_min(self.native, skip_nulls=True)
if not reverse
else pc.cumulative_min(self.native[::-1], skip_nulls=True)[::-1]
)
return self._with_native(result)
def cum_max(self, *, reverse: bool) -> Self:
result = (
pc.cumulative_max(self.native, skip_nulls=True)
if not reverse
else pc.cumulative_max(self.native[::-1], skip_nulls=True)[::-1]
)
return self._with_native(result)
def cum_prod(self, *, reverse: bool) -> Self:
result = (
pc.cumulative_prod(self.native, skip_nulls=True)
if not reverse
else pc.cumulative_prod(self.native[::-1], skip_nulls=True)[::-1]
)
return self._with_native(result)
def rolling_sum(self, window_size: int, *, min_samples: int, center: bool) -> Self:
min_samples = min_samples if min_samples is not None else window_size
padded_series, offset = pad_series(self, window_size=window_size, center=center)
cum_sum = padded_series.cum_sum(reverse=False).fill_null(
value=None, strategy="forward", limit=None
)
rolling_sum = (
cum_sum
- cum_sum.shift(window_size).fill_null(value=0, strategy=None, limit=None)
if window_size != 0
else cum_sum
)
valid_count = padded_series.cum_count(reverse=False)
count_in_window = valid_count - valid_count.shift(window_size).fill_null(
value=0, strategy=None, limit=None
)
result = self._with_native(
pc.if_else((count_in_window >= min_samples).native, rolling_sum.native, None)
)
return result._gather_slice(slice(offset, None))
def rolling_mean(self, window_size: int, *, min_samples: int, center: bool) -> Self:
min_samples = min_samples if min_samples is not None else window_size
padded_series, offset = pad_series(self, window_size=window_size, center=center)
cum_sum = padded_series.cum_sum(reverse=False).fill_null(
value=None, strategy="forward", limit=None
)
rolling_sum = (
cum_sum
- cum_sum.shift(window_size).fill_null(value=0, strategy=None, limit=None)
if window_size != 0
else cum_sum
)
valid_count = padded_series.cum_count(reverse=False)
count_in_window = valid_count - valid_count.shift(window_size).fill_null(
value=0, strategy=None, limit=None
)
result = (
self._with_native(
pc.if_else(
(count_in_window >= min_samples).native, rolling_sum.native, None
)
)
/ count_in_window
)
return result._gather_slice(slice(offset, None))
def rolling_var(
self, window_size: int, *, min_samples: int, center: bool, ddof: int
) -> Self:
min_samples = min_samples if min_samples is not None else window_size
padded_series, offset = pad_series(self, window_size=window_size, center=center)
cum_sum = padded_series.cum_sum(reverse=False).fill_null(
value=None, strategy="forward", limit=None
)
rolling_sum = (
cum_sum
- cum_sum.shift(window_size).fill_null(value=0, strategy=None, limit=None)
if window_size != 0
else cum_sum
)
cum_sum_sq = (
pow(padded_series, 2)
.cum_sum(reverse=False)
.fill_null(value=None, strategy="forward", limit=None)
)
rolling_sum_sq = (
cum_sum_sq
- cum_sum_sq.shift(window_size).fill_null(value=0, strategy=None, limit=None)
if window_size != 0
else cum_sum_sq
)
valid_count = padded_series.cum_count(reverse=False)
count_in_window = valid_count - valid_count.shift(window_size).fill_null(
value=0, strategy=None, limit=None
)
result = self._with_native(
pc.if_else(
(count_in_window >= min_samples).native,
(rolling_sum_sq - (rolling_sum**2 / count_in_window)).native,
None,
)
) / self._with_native(pc.max_element_wise((count_in_window - ddof).native, 0))
return result._gather_slice(slice(offset, None, None))
def rolling_std(
self, window_size: int, *, min_samples: int, center: bool, ddof: int
) -> Self:
return (
self.rolling_var(
window_size=window_size, min_samples=min_samples, center=center, ddof=ddof
)
** 0.5
)
def rank(self, method: RankMethod, *, descending: bool) -> Self:
if method == "average":
msg = (
"`rank` with `method='average' is not supported for pyarrow backend. "
"The available methods are {'min', 'max', 'dense', 'ordinal'}."
)
raise ValueError(msg)
sort_keys: Order = "descending" if descending else "ascending"
tiebreaker: TieBreaker = "first" if method == "ordinal" else method
native_series: ArrayOrChunkedArray
if self._backend_version < (14, 0, 0): # pragma: no cover
native_series = self.native.combine_chunks()
else:
native_series = self.native
null_mask = pc.is_null(native_series)
rank = pc.rank(native_series, sort_keys=sort_keys, tiebreaker=tiebreaker)
result = pc.if_else(null_mask, lit(None, rank.type), rank)
return self._with_native(result)
def hist_from_bins(
self, bins: list[float], *, include_breakpoint: bool
) -> ArrowDataFrame:
return (
_ArrowHist.from_series(self, include_breakpoint=include_breakpoint)
.with_bins(bins)
.to_frame()
)
def hist_from_bin_count(
self, bin_count: int, *, include_breakpoint: bool
) -> ArrowDataFrame:
return (
_ArrowHist.from_series(self, include_breakpoint=include_breakpoint)
.with_bin_count(bin_count)
.to_frame()
)
def __iter__(self) -> Iterator[Any]:
for x in self.native:
yield maybe_extract_py_scalar(x, return_py_scalar=True)
def __contains__(self, other: Any) -> bool:
from pyarrow import (
ArrowInvalid, # ignore-banned-imports
ArrowNotImplementedError, # ignore-banned-imports
ArrowTypeError, # ignore-banned-imports
)
try:
other_ = lit(other) if other is not None else lit(None, type=self._type)
return maybe_extract_py_scalar(
pc.is_in(other_, self.native), return_py_scalar=True
)
except (ArrowInvalid, ArrowNotImplementedError, ArrowTypeError) as exc:
msg = f"Unable to compare other of type {type(other)} with series of type {self.dtype}."
raise InvalidOperationError(msg) from exc
def log(self, base: float) -> Self:
return self._with_native(pc.logb(self.native, lit(base)))
def exp(self) -> Self:
return self._with_native(pc.exp(self.native))
def sqrt(self) -> Self:
return self._with_native(pc.sqrt(self.native))
@property
def dt(self) -> ArrowSeriesDateTimeNamespace:
return ArrowSeriesDateTimeNamespace(self)
@property
def cat(self) -> ArrowSeriesCatNamespace:
return ArrowSeriesCatNamespace(self)
@property
def str(self) -> ArrowSeriesStringNamespace:
return ArrowSeriesStringNamespace(self)
@property
def list(self) -> ArrowSeriesListNamespace:
return ArrowSeriesListNamespace(self)
@property
def struct(self) -> ArrowSeriesStructNamespace:
return ArrowSeriesStructNamespace(self)
ewm_mean = not_implemented()
class _ArrowHist(
EagerSeriesHist["ChunkedArrayAny", "list[ScalarAny] | pa.Int64Array | list[float]"]
):
_series: ArrowSeries
def to_frame(self) -> ArrowDataFrame:
# NOTE: Constructor typing is too strict for `TypedDict`
table: Incomplete = pa.Table.from_pydict
from_native = self._series.__narwhals_namespace__()._dataframe.from_native
return from_native(table(self._data), context=self._series)
# NOTE: *Could* be handled at narwhals-level
def is_empty_series(self) -> bool:
# NOTE: `ChunkedArray.combine_chunks` returns the concrete array type
# Stubs say `Array[pa.BooleanScalar]`, which is missing properties
# https://github.com/zen-xu/pyarrow-stubs/blob/6bedee748bc74feb8513b24bf43d64b24c7fddc8/pyarrow-stubs/__lib_pxi/array.pyi#L2395-L2399
is_null = self.native.is_null(nan_is_null=True)
arr = cast("pa.BooleanArray", is_null.combine_chunks())
return arr.false_count == 0
# NOTE: *Could* be handled at narwhals-level, **iff** we add `nw.repeat`, `nw.linear_space`
# See https://github.com/narwhals-dev/narwhals/pull/2839#discussion_r2215630696
def series_empty(self, arg: int | list[float], /) -> ArrowHistData:
count = self._zeros(arg)
if self._breakpoint:
return {"breakpoint": self._calculate_breakpoint(arg), "count": count}
return {"count": count}
def _zeros(self, arg: int | list[float], /) -> pa.Int64Array:
return zeros(arg) if isinstance(arg, int) else zeros(len(arg) - 1)
def _linear_space(
self,
start: float,
end: float,
num_samples: int,
*,
closed: Literal["both", "none"] = "both",
) -> _1DArray:
from numpy import linspace # ignore-banned-import
return linspace(start=start, stop=end, num=num_samples, endpoint=closed == "both")
def _calculate_bins(self, bin_count: int) -> _1DArray:
"""Prepare bins for histogram calculation from bin_count."""
d = pc.min_max(self.native)
lower, upper = d["min"].as_py(), d["max"].as_py()
if lower == upper:
lower -= 0.5
upper += 0.5
return self._linear_space(lower, upper, bin_count + 1)
def _calculate_hist(self, bins: list[float] | _1DArray) -> ArrowHistData:
ser = self.native
# NOTE: `mypy` refuses to resolve `ndarray.__getitem__`
# Previously annotated as `list[float]`, but
# - wasn't accurate to how we implemented it
# - `pa.scalar` overloads fail to match on `float | np.float64` (but runtime is fine)
bins = cast("list[float]", bins)
# Handle single bin case
if len(bins) == 2:
is_between_bins = pc.and_(
pc.greater_equal(ser, lit(bins[0])), pc.less_equal(ser, lit(bins[1]))
)
count = pc.sum(is_between_bins.cast(pa.uint8()))
if self._breakpoint:
return {"breakpoint": [bins[-1]], "count": [count]}
return {"count": [count]}
# Handle multiple bins
import numpy as np # ignore-banned-import
bin_indices = np.searchsorted(bins, ser, side="left")
# lowest bin is inclusive
bin_indices = pc.if_else(pc.equal(ser, lit(bins[0])), 1, bin_indices)
# Align unique categories and counts appropriately
obs_cats, obs_counts = np.unique(bin_indices, return_counts=True)
obj_cats = np.arange(1, len(bins))
counts = np.zeros_like(obj_cats)
counts[np.isin(obj_cats, obs_cats)] = obs_counts[np.isin(obs_cats, obj_cats)]
if self._breakpoint:
return {"breakpoint": bins[1:], "count": counts}
return {"count": counts}