PERF/REF: improve performance of Series.searchsorted, PandasArray.searchsorted, collect functionality (pandas-dev#22034)

Author: topper-123

asv_bench/benchmarks/series_methods.py

@@ -124,6 +124,25 @@ def time_dropna(self, dtype):
         self.s.dropna()
 
 
+class SearchSorted(object):
+
+    goal_time = 0.2
+    params = ['int8', 'int16', 'int32', 'int64',
+              'uint8', 'uint16', 'uint32', 'uint64',
+              'float16', 'float32', 'float64',
+              'str']
+    param_names = ['dtype']
+
+    def setup(self, dtype):
+        N = 10**5
+        data = np.array([1] * N + [2] * N + [3] * N).astype(dtype)
+        self.s = Series(data)
+
+    def time_searchsorted(self, dtype):
+        key = '2' if dtype == 'str' else 2
+        self.s.searchsorted(key)
+
+
 class Map(object):
 
     params = ['dict', 'Series']

doc/source/whatsnew/v0.25.0.rst

@@ -96,7 +96,8 @@ Performance Improvements
 
 - Significant speedup in `SparseArray` initialization that benefits most operations, fixing performance regression introduced in v0.20.0 (:issue:`24985`)
 - `DataFrame.to_stata()` is now faster when outputting data with any string or non-native endian columns (:issue:`25045`)
--
+- Improved performance of :meth:`Series.searchsorted`. The speedup is especially large when the dtype is
+  int8/int16/int32 and the searched key is within the integer bounds for the dtype (:issue:`22034`)
 
 
 .. _whatsnew_0250.bug_fixes:

pandas/core/algorithms.py

@@ -19,7 +19,7 @@
     ensure_float64, ensure_int64, ensure_object, ensure_platform_int,
     ensure_uint64, is_array_like, is_bool_dtype, is_categorical_dtype,
     is_complex_dtype, is_datetime64_any_dtype, is_datetime64tz_dtype,
-    is_datetimelike, is_extension_array_dtype, is_float_dtype,
+    is_datetimelike, is_extension_array_dtype, is_float_dtype, is_integer,
     is_integer_dtype, is_interval_dtype, is_list_like, is_numeric_dtype,
     is_object_dtype, is_period_dtype, is_scalar, is_signed_integer_dtype,
     is_sparse, is_timedelta64_dtype, is_unsigned_integer_dtype,
@@ -1724,6 +1724,89 @@ def func(arr, indexer, out, fill_value=np.nan):
     return out
 
 
+# ------------ #
+# searchsorted #
+# ------------ #
+
+def searchsorted(arr, value, side="left", sorter=None):
+    """
+    Find indices where elements should be inserted to maintain order.
+
+    .. versionadded:: 0.25.0
+
+    Find the indices into a sorted array `arr` (a) such that, if the
+    corresponding elements in `value` were inserted before the indices,
+    the order of `arr` would be preserved.
+
+    Assuming that `arr` is sorted:
+
+    ======  ================================
+    `side`  returned index `i` satisfies
+    ======  ================================
+    left    ``arr[i-1] < value <= self[i]``
+    right   ``arr[i-1] <= value < self[i]``
+    ======  ================================
+
+    Parameters
+    ----------
+    arr: array-like
+        Input array. If `sorter` is None, then it must be sorted in
+        ascending order, otherwise `sorter` must be an array of indices
+        that sort it.
+    value : array_like
+        Values to insert into `arr`.
+    side : {'left', 'right'}, optional
+        If 'left', the index of the first suitable location found is given.
+        If 'right', return the last such index.  If there is no suitable
+        index, return either 0 or N (where N is the length of `self`).
+    sorter : 1-D array_like, optional
+        Optional array of integer indices that sort array a into ascending
+        order. They are typically the result of argsort.
+
+    Returns
+    -------
+    array of ints
+        Array of insertion points with the same shape as `value`.
+
+    See Also
+    --------
+    numpy.searchsorted : Similar method from NumPy.
+    """
+    if sorter is not None:
+        sorter = ensure_platform_int(sorter)
+
+    if isinstance(arr, np.ndarray) and is_integer_dtype(arr) and (
+            is_integer(value) or is_integer_dtype(value)):
+        from .arrays.array_ import array
+        # if `arr` and `value` have different dtypes, `arr` would be
+        # recast by numpy, causing a slow search.
+        # Before searching below, we therefore try to give `value` the
+        # same dtype as `arr`, while guarding against integer overflows.
+        iinfo = np.iinfo(arr.dtype.type)
+        value_arr = np.array([value]) if is_scalar(value) else np.array(value)
+        if (value_arr >= iinfo.min).all() and (value_arr <= iinfo.max).all():
+            # value within bounds, so no overflow, so can convert value dtype
+            # to dtype of arr
+            dtype = arr.dtype
+        else:
+            dtype = value_arr.dtype
+
+        if is_scalar(value):
+            value = dtype.type(value)
+        else:
+            value = array(value, dtype=dtype)
+    elif not (is_object_dtype(arr) or is_numeric_dtype(arr) or
+              is_categorical_dtype(arr)):
+        from pandas.core.series import Series
+        # E.g. if `arr` is an array with dtype='datetime64[ns]'
+        # and `value` is a pd.Timestamp, we may need to convert value
+        value_ser = Series(value)._values
+        value = value_ser[0] if is_scalar(value) else value_ser
+
+    result = arr.searchsorted(value, side=side, sorter=sorter)
+    return result
+
+
 # ---- #
 # diff #
 # ---- #

pandas/core/arrays/base.py

@@ -555,17 +555,17 @@ def searchsorted(self, value, side="left", sorter=None):
         .. versionadded:: 0.24.0
 
         Find the indices into a sorted array `self` (a) such that, if the
-        corresponding elements in `v` were inserted before the indices, the
-        order of `self` would be preserved.
+        corresponding elements in `value` were inserted before the indices,
+        the order of `self` would be preserved.
 
-        Assuming that `a` is sorted:
+        Assuming that `self` is sorted:
 
-        ======  ============================
+        ======  ================================
         `side`  returned index `i` satisfies
-        ======  ============================
-        left    ``self[i-1] < v <= self[i]``
-        right   ``self[i-1] <= v < self[i]``
-        ======  ============================
+        ======  ================================
+        left    ``self[i-1] < value <= self[i]``
+        right   ``self[i-1] <= value < self[i]``
+        ======  ================================
 
         Parameters
         ----------
@@ -581,7 +581,7 @@ def searchsorted(self, value, side="left", sorter=None):
 
         Returns
         -------
-        indices : array of ints
+        array of ints
             Array of insertion points with the same shape as `value`.
 
         See Also

pandas/core/arrays/numpy_.py

@@ -4,6 +4,7 @@
 
 from pandas._libs import lib
 from pandas.compat.numpy import function as nv
+from pandas.util._decorators import Appender
 from pandas.util._validators import validate_fillna_kwargs
 
 from pandas.core.dtypes.dtypes import ExtensionDtype
@@ -12,6 +13,7 @@
 
 from pandas import compat
 from pandas.core import nanops
+from pandas.core.algorithms import searchsorted
 from pandas.core.missing import backfill_1d, pad_1d
 
 from .base import ExtensionArray, ExtensionOpsMixin
@@ -423,6 +425,11 @@ def to_numpy(self, dtype=None, copy=False):
 
         return result
 
+    @Appender(ExtensionArray.searchsorted.__doc__)
+    def searchsorted(self, value, side='left', sorter=None):
+        return searchsorted(self.to_numpy(), value,
+                            side=side, sorter=sorter)
+
     # ------------------------------------------------------------------------
     # Ops
 

pandas/core/base.py

@@ -1522,11 +1522,11 @@ def factorize(self, sort=False, na_sentinel=-1):
         array([3])
         """)
 
-    @Substitution(klass='IndexOpsMixin')
+    @Substitution(klass='Index')
     @Appender(_shared_docs['searchsorted'])
     def searchsorted(self, value, side='left', sorter=None):
-        # needs coercion on the key (DatetimeIndex does already)
-        return self._values.searchsorted(value, side=side, sorter=sorter)
+        return algorithms.searchsorted(self._values, value,
+                                       side=side, sorter=sorter)
 
     def drop_duplicates(self, keep='first', inplace=False):
         inplace = validate_bool_kwarg(inplace, 'inplace')

pandas/core/series.py

@@ -2392,12 +2392,8 @@ def __rmatmul__(self, other):
     @Substitution(klass='Series')
     @Appender(base._shared_docs['searchsorted'])
     def searchsorted(self, value, side='left', sorter=None):
-        if sorter is not None:
-            sorter = ensure_platform_int(sorter)
-        result = self._values.searchsorted(Series(value)._values,
-                                           side=side, sorter=sorter)
-
-        return result[0] if is_scalar(value) else result
+        return algorithms.searchsorted(self._values, value,
+                                       side=side, sorter=sorter)
 
     # -------------------------------------------------------------------
     # Combination

pandas/tests/arrays/test_array.py

@@ -9,6 +9,7 @@
 
 import pandas as pd
 from pandas.api.extensions import register_extension_dtype
+from pandas.api.types import is_scalar
 from pandas.core.arrays import PandasArray, integer_array, period_array
 from pandas.tests.extension.decimal import (
     DecimalArray, DecimalDtype, to_decimal)
@@ -254,3 +255,51 @@ def test_array_not_registered(registry_without_decimal):
     result = pd.array(data, dtype=DecimalDtype)
     expected = DecimalArray._from_sequence(data)
     tm.assert_equal(result, expected)
+
+
+class TestArrayAnalytics(object):
+    def test_searchsorted(self, string_dtype):
+        arr = pd.array(['a', 'b', 'c'], dtype=string_dtype)
+
+        result = arr.searchsorted('a', side='left')
+        assert is_scalar(result)
+        assert result == 0
+
+        result = arr.searchsorted('a', side='right')
+        assert is_scalar(result)
+        assert result == 1
+
+    def test_searchsorted_numeric_dtypes_scalar(self, any_real_dtype):
+        arr = pd.array([1, 3, 90], dtype=any_real_dtype)
+        result = arr.searchsorted(30)
+        assert is_scalar(result)
+        assert result == 2
+
+        result = arr.searchsorted([30])
+        expected = np.array([2], dtype=np.intp)
+        tm.assert_numpy_array_equal(result, expected)
+
+    def test_searchsorted_numeric_dtypes_vector(self, any_real_dtype):
+        arr = pd.array([1, 3, 90], dtype=any_real_dtype)
+        result = arr.searchsorted([2, 30])
+        expected = np.array([1, 2], dtype=np.intp)
+        tm.assert_numpy_array_equal(result, expected)
+
+    @pytest.mark.parametrize('arr, val', [
+        [pd.date_range('20120101', periods=10, freq='2D'),
+         pd.Timestamp('20120102')],
+        [pd.date_range('20120101', periods=10, freq='2D', tz='Asia/Hong_Kong'),
+         pd.Timestamp('20120102', tz='Asia/Hong_Kong')],
+        [pd.timedelta_range(start='1 day', end='10 days', periods=10),
+         pd.Timedelta('2 days')]])
+    def test_search_sorted_datetime64_scalar(self, arr, val):
+        arr = pd.array(arr)
+        result = arr.searchsorted(val)
+        assert is_scalar(result)
+        assert result == 1
+
+    def test_searchsorted_sorter(self, any_real_dtype):
+        arr = pd.array([3, 1, 2], dtype=any_real_dtype)
+        result = arr.searchsorted([0, 3], sorter=np.argsort(arr))
+        expected = np.array([0, 2], dtype=np.intp)
+        tm.assert_numpy_array_equal(result, expected)