PERF: Cythonize groupby transforms #4095

Author: chris-b1

asv_bench/benchmarks/groupby.py

@@ -772,3 +772,19 @@ def setup(self):
 
     def time_groupby_transform_series2(self):
         self.df.groupby('id')['val'].transform(np.mean)
+
+class groupby_transform_cythonized(object):
+    goal_time = 0.2
+
+    def setup(self):
+        np.random.seed(0)
+        self.df = DataFrame({'id': (np.arange(100000) / 3), 'val': np.random.randn(100000), })
+
+    def time_groupby_transform_cumprod(self):
+        self.df.groupby('id').cumprod()
+
+    def time_groupby_transform_cumsum(self):
+        self.df.groupby('id').cumsum()
+
+    def time_groupby_transform_shift(self):
+        self.df.groupby('id').shift()

doc/source/whatsnew/v0.17.1.txt

@@ -93,6 +93,8 @@ Performance Improvements
 - Performance bug in repr of ``Categorical`` categories, which was rendering the strings before chopping them for display (:issue:`11305`)
 - Improved performance of ``Series`` constructor with no data and ``DatetimeIndex`` (:issue:`11433`)
 
+- Improved performance ``shift``, ``cumprod``, and ``cumsum`` with groupby (:issue:`4095`)
+
 .. _whatsnew_0171.bug_fixes:
 
 

pandas/algos.pyx

@@ -50,7 +50,6 @@ cdef np.float64_t MAXfloat64 = np.inf
 cdef double NaN = <double> np.NaN
 cdef double nan = NaN
 
-
 cdef inline int int_max(int a, int b): return a if a >= b else b
 cdef inline int int_min(int a, int b): return a if a <= b else b
 
@@ -2266,43 +2265,6 @@ def group_last_bin_object(ndarray[object, ndim=2] out,
             else:
                 out[i, j] = resx[i, j]
 
-
-
-#----------------------------------------------------------------------
-# median
-
-def group_median(ndarray[float64_t, ndim=2] out,
-                 ndarray[int64_t] counts,
-                 ndarray[float64_t, ndim=2] values,
-                 ndarray[int64_t] labels):
-    '''
-    Only aggregates on axis=0
-    '''
-    cdef:
-        Py_ssize_t i, j, N, K, ngroups, size
-        ndarray[int64_t] _counts
-        ndarray data
-        float64_t* ptr
-    ngroups = len(counts)
-    N, K = (<object> values).shape
-
-    indexer, _counts = groupsort_indexer(labels, ngroups)
-    counts[:] = _counts[1:]
-
-    data = np.empty((K, N), dtype=np.float64)
-    ptr = <float64_t*> data.data
-
-    take_2d_axis1_float64_float64(values.T, indexer, out=data)
-
-    for i in range(K):
-        # exclude NA group
-        ptr += _counts[0]
-        for j in range(ngroups):
-            size = _counts[j + 1]
-            out[j, i] = _median_linear(ptr, size)
-            ptr += size
-
-
 cdef inline float64_t _median_linear(float64_t* a, int n):
     cdef int i, j, na_count = 0
     cdef float64_t result