PERF-#7657: Fork pandas eval and query implementation to improve performance.

modin/core/computation/align.py

@@ -0,0 +1,199 @@
+"""
+Core eval alignment algorithms. Forked from pandas.core.computation.align
+"""
+
+from __future__ import annotations
+
+import warnings
+from collections.abc import Sequence
+from functools import (
+    partial,
+    wraps,
+)
+from typing import (
+    Callable,
+)
+
+import numpy as np
+import pandas
+import pandas.core.common as com
+from pandas._typing import F
+from pandas.core.base import PandasObject
+from pandas.errors import PerformanceWarning
+
+from modin.core.computation.common import result_type_many
+from modin.pandas import DataFrame, Series
+from modin.pandas.base import BasePandasDataset
+
+
+def _align_core_single_unary_op(
+    term,
+) -> tuple[partial | type[BasePandasDataset], dict[str, pandas.Index] | None]:
+    typ: partial | type[BasePandasDataset]
+    axes: dict[str, pandas.Index] | None = None
+
+    if isinstance(term.value, np.ndarray):
+        typ = partial(np.asanyarray, dtype=term.value.dtype)
+    else:
+        typ = type(term.value)
+        if hasattr(term.value, "axes"):
+            axes = _zip_axes_from_type(typ, term.value.axes)
+
+    return typ, axes
+
+
+def _zip_axes_from_type(
+    typ: type[BasePandasDataset], new_axes: Sequence[pandas.Index]
+) -> dict[str, pandas.Index]:
+    return {name: new_axes[i] for i, name in enumerate(typ._AXIS_ORDERS)}
+
+
+def _any_pandas_objects(terms) -> bool:
+    """
+    Check a sequence of terms for instances of PandasObject.
+    """
+    return any(isinstance(term.value, PandasObject) for term in terms)
+
+
+def _filter_special_cases(f) -> Callable[[F], F]:
+    @wraps(f)
+    def wrapper(terms):
+        # single unary operand
+        if len(terms) == 1:
+            return _align_core_single_unary_op(terms[0])
+
+        term_values = (term.value for term in terms)
+
+        # we don't have any pandas objects
+        if not _any_pandas_objects(terms):
+            return result_type_many(*term_values), None
+
+        return f(terms)
+
+    return wrapper
+
+
+@_filter_special_cases
+def _align_core(terms):
+    term_index = [i for i, term in enumerate(terms) if hasattr(term.value, "axes")]
+    term_dims = [terms[i].value.ndim for i in term_index]
+
+    ndims = pandas.Series(dict(zip(term_index, term_dims)))
+
+    # initial axes are the axes of the largest-axis'd term
+    biggest = terms[ndims.idxmax()].value
+    typ = biggest._constructor
+    axes = biggest.axes
+    naxes = len(axes)
+    gt_than_one_axis = naxes > 1
+
+    for value in (terms[i].value for i in term_index):
+        is_series = isinstance(value, Series)
+        is_series_and_gt_one_axis = is_series and gt_than_one_axis
+
+        for axis, items in enumerate(value.axes):
+            if is_series_and_gt_one_axis:
+                ax, itm = naxes - 1, value.index
+            else:
+                ax, itm = axis, items
+
+            if not axes[ax].is_(itm):
+                axes[ax] = axes[ax].union(itm)
+
+    for i, ndim in ndims.items():
+        for axis, items in zip(range(ndim), axes):
+            ti = terms[i].value
+
+            if hasattr(ti, "reindex"):
+                transpose = isinstance(ti, Series) and naxes > 1
+                reindexer = axes[naxes - 1] if transpose else items
+
+                term_axis_size = len(ti.axes[axis])
+                reindexer_size = len(reindexer)
+
+                ordm = np.log10(max(1, abs(reindexer_size - term_axis_size)))
+                if ordm >= 1 and reindexer_size >= 10000:
+                    w = (
+                        f"Alignment difference on axis {axis} is larger "
+                        + f"than an order of magnitude on term {repr(terms[i].name)}, "
+                        + f"by more than {ordm:.4g}; performance may suffer."
+                    )
+                    warnings.warn(w, category=PerformanceWarning)
+
+                obj = ti.reindex(reindexer, axis=axis, copy=False)
+                terms[i].update(obj)
+
+        terms[i].update(terms[i].value.values)
+
+    return typ, _zip_axes_from_type(typ, axes)
+
+
+def align_terms(terms):
+    """
+    Align a set of terms.
+    """
+    try:
+        # flatten the parse tree (a nested list, really)
+        terms = list(com.flatten(terms))
+    except TypeError:
+        # can't iterate so it must just be a constant or single variable
+        if isinstance(terms.value, (Series, DataFrame)):
+            typ = type(terms.value)
+            return typ, _zip_axes_from_type(typ, terms.value.axes)
+        return np.result_type(terms.type), None
+
+    # if all resolved variables are numeric scalars
+    if all(term.is_scalar for term in terms):
+        return result_type_many(*(term.value for term in terms)).type, None
+
+    # perform the main alignment
+    typ, axes = _align_core(terms)
+    return typ, axes
+
+
+def reconstruct_object(typ, obj, axes, dtype):
+    """
+    Reconstruct an object given its type, raw value, and possibly empty
+    (None) axes.
+
+    Parameters
+    ----------
+    typ : object
+        A type
+    obj : object
+        The value to use in the type constructor
+    axes : dict
+        The axes to use to construct the resulting pandas object
+
+    Returns
+    -------
+    ret : typ
+        An object of type ``typ`` with the value `obj` and possible axes
+        `axes`.
+    """
+    try:
+        typ = typ.type
+    except AttributeError:
+        pass
+
+    res_t = np.result_type(obj.dtype, dtype)
+
+    if not isinstance(typ, partial) and issubclass(typ, PandasObject):
+        return typ(obj, dtype=res_t, **axes)
+
+    # special case for pathological things like ~True/~False
+    if hasattr(res_t, "type") and typ == np.bool_ and res_t != np.bool_:
+        ret_value = res_t.type(obj)
+    else:
+        ret_value = typ(obj).astype(res_t)
+        # The condition is to distinguish 0-dim array (returned in case of
+        # scalar) and 1 element array
+        # e.g. np.array(0) and np.array([0])
+        if (
+            len(obj.shape) == 1
+            and len(obj) == 1
+            and not isinstance(ret_value, np.ndarray)
+        ):
+            ret_value = np.array([ret_value]).astype(res_t)
+
+    return ret_value

modin/core/computation/check.py

@@ -0,0 +1,12 @@
+"""
+Forked from pandas.core.computation.check
+"""
+
+from __future__ import annotations
+
+from pandas.compat._optional import import_optional_dependency
+
+ne = import_optional_dependency("numexpr", errors="warn")
+NUMEXPR_INSTALLED = ne is not None
+
+__all__ = ["NUMEXPR_INSTALLED"]

modin/core/computation/common.py

@@ -0,0 +1,51 @@
+"""
+Forked from pandas.core.computation.common
+"""
+
+from __future__ import annotations
+
+from functools import reduce
+
+import numpy as np
+from pandas._config import get_option
+
+
+def ensure_decoded(s) -> str:
+    """
+    If we have bytes, decode them to unicode.
+    """
+    if isinstance(s, (np.bytes_, bytes)):
+        s = s.decode(get_option("display.encoding"))
+    return s
+
+
+def result_type_many(*arrays_and_dtypes):
+    """
+    Wrapper around numpy.result_type which overcomes the NPY_MAXARGS (32)
+    argument limit.
+    """
+    try:
+        return np.result_type(*arrays_and_dtypes)
+    except ValueError:
+        # we have > NPY_MAXARGS terms in our expression
+        return reduce(np.result_type, arrays_and_dtypes)
+    except TypeError:
+        from pandas.core.dtypes.cast import find_common_type
+        from pandas.core.dtypes.common import is_extension_array_dtype
+
+        arr_and_dtypes = list(arrays_and_dtypes)
+        ea_dtypes, non_ea_dtypes = [], []
+        for arr_or_dtype in arr_and_dtypes:
+            if is_extension_array_dtype(arr_or_dtype):
+                ea_dtypes.append(arr_or_dtype)
+            else:
+                non_ea_dtypes.append(arr_or_dtype)
+
+        if non_ea_dtypes:
+            try:
+                np_dtype = np.result_type(*non_ea_dtypes)
+            except ValueError:
+                np_dtype = reduce(np.result_type, arrays_and_dtypes)
+            return find_common_type(ea_dtypes + [np_dtype])
+
+        return find_common_type(ea_dtypes)