BUG: datetime rolling min/max segfaults when closed=left (pandas-dev#21704)

User reported that `df.rolling(to_offset('3D'), closed='left').max()`
segfaults when df has a datetime index. The bug was in PR pandas-dev#19549. In
that PR, in https://github.com/pandas-dev/pandas/blame/master/pandas/_libs/window.pyx#L1268
`i` is initialized to `endi[0]`, which is 0 when `closed=left`.
So in the next line when it tries to set `output[i-1]` it goes out of bounds.
In addition, there are 2 more bugs in the `roll_min_max` code.
The second bug is that for variable size windows, the `nobs` is never updated
when elements leave the window. The third bug is at the end of the fixed
window where all output elements up to `minp` are initialized to 0 if
the input is not float.

This PR fixes all three of the aforementioned bugs, at the cost of casting the
output array to floating point even if the input is integer. This is less
than ideal if the output has no NaNs, but is still consistent with roll_sum
behavior.

Author: changhiskhan

doc/source/whatsnew/v0.24.0.txt

@@ -398,6 +398,9 @@ Groupby/Resample/Rolling
 -
 -
 
+- Multiple bugs in :func:`pandas.core.Rolling.min` with ``closed='left'` and a
+  datetime-like index leading to incorrect results and also segfault. (:issue:`21704`)
+
 Sparse
 ^^^^^^
 

pandas/_libs/window.pyx

@@ -1222,11 +1222,12 @@ cdef _roll_min_max(ndarray[numeric] input, int64_t win, int64_t minp,
     cdef:
         numeric ai
         bint is_variable, should_replace
-        int64_t N, i, removed, window_i
+        int64_t N, i, removed, window_i, close_offset
         Py_ssize_t nobs = 0
         deque Q[int64_t]
+        deque W[int64_t]
         ndarray[int64_t] starti, endi
-        ndarray[numeric, ndim=1] output
+        ndarray[double_t, ndim=1] output
     cdef:
         int64_t* death
         numeric* ring
@@ -1241,49 +1242,69 @@ cdef _roll_min_max(ndarray[numeric] input, int64_t win, int64_t minp,
         input, win,
         minp, index, closed)
 
-    output = np.empty(N, dtype=input.dtype)
+    output = np.empty(N, dtype=float)
 
+    # The front should always be the minimum in the window
     Q = deque[int64_t]()
 
     if is_variable:
-
+        # need this to maintain the window for nobs calc
+        W = deque[int64_t]()
         with nogil:
 
             # This is using a modified version of the C++ code in this
             # SO post: http://bit.ly/2nOoHlY
             # The original impl didn't deal with variable window sizes
             # So the code was optimized for that
 
-            for i from starti[0] <= i < endi[0]:
-                ai = init_mm(input[i], &nobs, is_max)
-
-                if is_max:
-                    while not Q.empty() and ai >= input[Q.back()]:
-                        Q.pop_back()
-                else:
-                    while not Q.empty() and ai <= input[Q.back()]:
-                        Q.pop_back()
-                Q.push_back(i)
-
-            for i from endi[0] <= i < N:
-                output[i-1] = calc_mm(minp, nobs, input[Q.front()])
+            if endi[0] > starti[0]:
+                # right is closed
+                for i from starti[0] <= i < endi[0]:
+                    ai = init_mm(input[i], &nobs, is_max)
 
+                    if is_max:
+                        while not Q.empty() and ai >= input[Q.back()]:
+                            Q.pop_back()
+                    else:
+                        while not Q.empty() and ai <= input[Q.back()]:
+                            Q.pop_back()
+                    Q.push_back(i)
+                    W.push_back(i)
+                output[0] = calc_mm(minp, nobs, input[Q.front()])
+                close_offset = 0
+            else:
+                init_mm(input[endi[0]], &nobs, is_max)
+                Q.push_back(endi[0])
+                W.push_back(endi[0])
+                # if right is open then the first output is always NaN
+                output[0] = NaN
+                close_offset = 1
+
+            for i from endi[close_offset] <= i < N:
                 ai = init_mm(input[i], &nobs, is_max)
 
+                # Discard previous entries if we find new min or max
                 if is_max:
                     while not Q.empty() and ai >= input[Q.back()]:
                         Q.pop_back()
                 else:
                     while not Q.empty() and ai <= input[Q.back()]:
                         Q.pop_back()
 
-                while not Q.empty() and Q.front() <= i - (endi[i] - starti[i]):
+                # Remove elements leaving the window
+                while not Q.empty() and Q.front() <= (i - (endi[i] - starti[i])
+                                                      - close_offset):
                     Q.pop_front()
 
-                Q.push_back(i)
-
-            output[N-1] = calc_mm(minp, nobs, input[Q.front()])
+                # Update the nobs
+                while not W.empty() and W.front() <= (i - (endi[i] - starti[i])
+                                                      - close_offset):
+                    remove_mm(input[W.front()], &nobs)
+                    W.pop_front()
 
+                Q.push_back(i)
+                W.push_back(i)
+                output[i] = calc_mm(minp, nobs, input[Q.front()])
     else:
         # setup the rings of death!
         ring = <numeric *>malloc(win * sizeof(numeric))

pandas/tests/test_window.py

@@ -464,6 +464,49 @@ def test_closed(self):
         with pytest.raises(ValueError):
             df.rolling(window=3, closed='neither')
 
+    def test_closed_min_max_datetime(self):
+        # see gh-21704
+        ser = pd.Series(data=np.arange(10),
+                        index=pd.date_range('2000', periods=10))
+
+        result = ser.rolling('3D', closed='right').min()
+        expected = pd.Series([0.0, 0, 0, 1, 2, 3, 4, 5, 6, 7],
+                             index=ser.index)
+        tm.assert_series_equal(result, expected)
+
+        result = ser.rolling('3D', closed='both').min()
+        expected = pd.Series([0.0, 0, 0, 0, 1, 2, 3, 4, 5, 6],
+                             index=ser.index)
+        tm.assert_series_equal(result, expected)
+
+        result = ser.rolling('3D', closed='neither').min()
+        expected = pd.Series([np.nan, 0, 0, 1, 2, 3, 4, 5, 6, 7],
+                             index=ser.index)
+        tm.assert_series_equal(result, expected)
+
+        # shouldn't segfault
+        result = ser.rolling('3D', closed='left').min()
+        expected = pd.Series([np.nan, 0, 0, 0, 1, 2, 3, 4, 5, 6],
+                             index=ser.index)
+        tm.assert_series_equal(result, expected)
+
+        # uneven
+        ser = ser.drop(index=ser.index[[1, 5]])
+        result = ser.rolling('3D', closed='left').min()
+        expected = pd.Series([np.nan, 0, 0, 2, 3, 4, 6, 6],
+                             index=ser.index)
+        tm.assert_series_equal(result, expected)
+
+    def test_closed_min_max_minp(self):
+        # see gh-21704
+        ser = pd.Series(data=np.arange(10),
+                        index=pd.date_range('2000', periods=10))
+        ser[ser.index[-3:]] = np.nan
+        result = ser.rolling('3D', min_periods=2, closed='left').min()
+        expected = pd.Series([np.nan, 0, 0, 0, 1, 2, 3, 4, 5, np.nan],
+                             index=ser.index)
+        tm.assert_series_equal(result, expected)
+
     @pytest.mark.parametrize('roller', ['1s', 1])
     def tests_empty_df_rolling(self, roller):
         # GH 15819 Verifies that datetime and integer rolling windows can be