Backport PR #42338: PERF/REGR: revert #41785 (#42444)

Co-authored-by: jbrockmendel <[email protected]>

Author: meeseeksmachine

doc/source/whatsnew/v1.3.1.rst

@@ -16,6 +16,7 @@ Fixed regressions
 ~~~~~~~~~~~~~~~~~
 - Pandas could not be built on PyPy (:issue:`42355`)
 - :class:`DataFrame` constructed with with an older version of pandas could not be unpickled (:issue:`42345`)
+- Performance regression in constructing a :class:`DataFrame` from a dictionary of dictionaries (:issue:`42338`)
 -
 
 .. ---------------------------------------------------------------------------

pandas/_libs/lib.pyi

@@ -51,6 +51,7 @@ def is_string_array(values: np.ndarray, skipna: bool = False): ...
 def is_float_array(values: np.ndarray, skipna: bool = False): ...
 def is_integer_array(values: np.ndarray, skipna: bool = False): ...
 def is_bool_array(values: np.ndarray, skipna: bool = False): ...
+def fast_multiget(mapping: dict, keys: np.ndarray, default=np.nan) -> np.ndarray: ...
 def fast_unique_multiple_list_gen(gen: Generator, sort: bool = True) -> list: ...
 def fast_unique_multiple_list(lists: list, sort: bool = True) -> list: ...
 def fast_unique_multiple(arrays: list, sort: bool = True) -> list: ...

pandas/_libs/lib.pyx

@@ -2979,6 +2979,28 @@ def to_object_array_tuples(rows: object) -> np.ndarray:
     return result
 
 
+@cython.wraparound(False)
+@cython.boundscheck(False)
+def fast_multiget(dict mapping, ndarray keys, default=np.nan) -> np.ndarray:
+    cdef:
+        Py_ssize_t i, n = len(keys)
+        object val
+        ndarray[object] output = np.empty(n, dtype='O')
+
+    if n == 0:
+        # kludge, for Series
+        return np.empty(0, dtype='f8')
+
+    for i in range(n):
+        val = keys[i]
+        if val in mapping:
+            output[i] = mapping[val]
+        else:
+            output[i] = default
+
+    return maybe_convert_objects(output)
+
+
 def is_bool_list(obj: list) -> bool:
     """
     Check if this list contains only bool or np.bool_ objects.

pandas/core/dtypes/cast.py

@@ -781,6 +781,21 @@ def infer_dtype_from_scalar(val, pandas_dtype: bool = False) -> tuple[DtypeObj,
     return dtype, val
 
 
+def dict_compat(d: dict[Scalar, Scalar]) -> dict[Scalar, Scalar]:
+    """
+    Convert datetimelike-keyed dicts to a Timestamp-keyed dict.
+
+    Parameters
+    ----------
+    d: dict-like object
+
+    Returns
+    -------
+    dict
+    """
+    return {maybe_box_datetimelike(key): value for key, value in d.items()}
+
+
 def infer_dtype_from_array(
     arr, pandas_dtype: bool = False
 ) -> tuple[DtypeObj, ArrayLike]:

pandas/core/internals/construction.py

@@ -26,6 +26,7 @@
 
 from pandas.core.dtypes.cast import (
     construct_1d_arraylike_from_scalar,
+    dict_compat,
     maybe_cast_to_datetime,
     maybe_convert_platform,
     maybe_infer_to_datetimelike,
@@ -59,15 +60,16 @@
     TimedeltaArray,
 )
 from pandas.core.construction import (
-    create_series_with_explicit_dtype,
     ensure_wrapped_if_datetimelike,
     extract_array,
     range_to_ndarray,
     sanitize_array,
 )
 from pandas.core.indexes import base as ibase
 from pandas.core.indexes.api import (
+    DatetimeIndex,
     Index,
+    TimedeltaIndex,
     ensure_index,
     get_objs_combined_axis,
     union_indexes,
@@ -556,6 +558,7 @@ def convert(v):
 
 
 def _homogenize(data, index: Index, dtype: DtypeObj | None) -> list[ArrayLike]:
+    oindex = None
     homogenized = []
 
     for val in data:
@@ -570,9 +573,18 @@ def _homogenize(data, index: Index, dtype: DtypeObj | None) -> list[ArrayLike]:
             val = val._values
         else:
             if isinstance(val, dict):
-                # see test_constructor_subclass_dict
-                #  test_constructor_dict_datetime64_index
-                val = create_series_with_explicit_dtype(val, index=index)._values
+                # GH#41785 this _should_ be equivalent to (but faster than)
+                #  val = create_series_with_explicit_dtype(val, index=index)._values
+                if oindex is None:
+                    oindex = index.astype("O")
+
+                if isinstance(index, (DatetimeIndex, TimedeltaIndex)):
+                    # see test_constructor_dict_datetime64_index
+                    val = dict_compat(val)
+                else:
+                    # see test_constructor_subclass_dict
+                    val = dict(val)
+                val = lib.fast_multiget(val, oindex._values, default=np.nan)
 
             val = sanitize_array(
                 val, index, dtype=dtype, copy=False, raise_cast_failure=False

pandas/tests/dtypes/cast/test_dict_compat.py

@@ -0,0 +1,14 @@
+import numpy as np
+
+from pandas.core.dtypes.cast import dict_compat
+
+from pandas import Timestamp
+
+
+def test_dict_compat():
+    data_datetime64 = {np.datetime64("1990-03-15"): 1, np.datetime64("2015-03-15"): 2}
+    data_unchanged = {1: 2, 3: 4, 5: 6}
+    expected = {Timestamp("1990-3-15"): 1, Timestamp("2015-03-15"): 2}
+    assert dict_compat(data_datetime64) == expected
+    assert dict_compat(expected) == expected
+    assert dict_compat(data_unchanged) == data_unchanged