PERF: performance regression in Series.asof (#14476)

* Fix performance regression in Series.asof by avoiding pre-computing nulls and returning value by indexing the underlying ndarray.

Author: laudney

asv_bench/benchmarks/timeseries.py

@@ -284,56 +284,77 @@ class timeseries_asof(object):
     goal_time = 0.2
 
     def setup(self):
-        self.N = 100000
-        self.rng = date_range(start='1/1/2000', periods=self.N, freq='T')
-        if hasattr(Series, 'convert'):
-            Series.resample = Series.convert
-        self.ts = Series(np.random.randn(self.N), index=self.rng)
         self.N = 10000
         self.rng = date_range(start='1/1/1990', periods=self.N, freq='53s')
-        self.ts = Series(np.random.randn(self.N), index=self.rng)
         self.dates = date_range(start='1/1/1990', periods=(self.N * 10), freq='5s')
+        self.ts = Series(np.random.randn(self.N), index=self.rng)
+        self.ts2 = self.ts.copy()
+        self.ts2[250:5000] = np.nan
+        self.ts3 = self.ts.copy()
+        self.ts3[-5000:] = np.nan
 
-    def time_timeseries_asof(self):
+    # test speed of pre-computing NAs.
+    def time_asof_list(self):
         self.ts.asof(self.dates)
 
+    # should be roughly the same as above.
+    def time_asof_nan_list(self):
+        self.ts2.asof(self.dates)
 
-class timeseries_asof_nan(object):
-    goal_time = 0.2
+    # test speed of the code path for a scalar index
+    # without *while* loop
+    def time_asof_single(self):
+        self.ts.asof(self.dates[0])
 
-    def setup(self):
-        self.N = 100000
-        self.rng = date_range(start='1/1/2000', periods=self.N, freq='T')
-        if hasattr(Series, 'convert'):
-            Series.resample = Series.convert
-        self.ts = Series(np.random.randn(self.N), index=self.rng)
-        self.N = 10000
-        self.rng = date_range(start='1/1/1990', periods=self.N, freq='53s')
-        self.ts = Series(np.random.randn(self.N), index=self.rng)
-        self.dates = date_range(start='1/1/1990', periods=(self.N * 10), freq='5s')
-        self.ts[250:5000] = np.nan
+    # test speed of the code path for a scalar index
+    # before the start. should be the same as above.
+    def time_asof_single_early(self):
+        self.ts.asof(self.dates[0] - dt.timedelta(10))
 
-    def time_timeseries_asof_nan(self):
-        self.ts.asof(self.dates)
+    # test the speed of the code path for a scalar index
+    # with a long *while* loop. should still be much
+    # faster than pre-computing all the NAs.
+    def time_asof_nan_single(self):
+        self.ts3.asof(self.dates[-1])
 
 
-class timeseries_asof_single(object):
+class timeseries_dataframe_asof(object):
     goal_time = 0.2
 
     def setup(self):
-        self.N = 100000
-        self.rng = date_range(start='1/1/2000', periods=self.N, freq='T')
-        if hasattr(Series, 'convert'):
-            Series.resample = Series.convert
-        self.ts = Series(np.random.randn(self.N), index=self.rng)
         self.N = 10000
+        self.M = 100
         self.rng = date_range(start='1/1/1990', periods=self.N, freq='53s')
-        self.ts = Series(np.random.randn(self.N), index=self.rng)
         self.dates = date_range(start='1/1/1990', periods=(self.N * 10), freq='5s')
+        self.ts = DataFrame(np.random.randn(self.N, self.M), index=self.rng)
+        self.ts2 = self.ts.copy()
+        self.ts2.iloc[250:5000] = np.nan
+        self.ts3 = self.ts.copy()
+        self.ts3.iloc[-5000:] = np.nan
+
+    # test speed of pre-computing NAs.
+    def time_asof_list(self):
+        self.ts.asof(self.dates)
 
-    def time_timeseries_asof_single(self):
+    # should be roughly the same as above.
+    def time_asof_nan_list(self):
+        self.ts2.asof(self.dates)
+
+    # test speed of the code path for a scalar index
+    # with pre-computing all NAs.
+    def time_asof_single(self):
         self.ts.asof(self.dates[0])
 
+    # should be roughly the same as above.
+    def time_asof_nan_single(self):
+        self.ts3.asof(self.dates[-1])
+
+    # test speed of the code path for a scalar index
+    # before the start. should be without the cost of
+    # pre-computing all the NAs.
+    def time_asof_single_early(self):
+        self.ts.asof(self.dates[0] - dt.timedelta(10))
+
 
 class timeseries_custom_bday_apply(object):
     goal_time = 0.2

doc/source/whatsnew/v0.19.1.txt

@@ -21,8 +21,9 @@ Performance Improvements
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
 - Fixed performance regression in factorization of ``Period`` data (:issue:`14338`)
-- Improved Performance in ``.to_json()`` when ``lines=True`` (:issue:`14408`)
-
+- Improved performance in ``.to_json()`` when ``lines=True`` (:issue:`14408`)
+- Improved performance in ``Series.asof(where)`` when ``where`` is a scalar (:issue:`14461)
+- Improved performance in ``DataFrame.asof(where)`` when ``where`` is a scalar (:issue:`14461)
 
 
 

pandas/core/generic.py

@@ -3735,10 +3735,10 @@ def asof(self, where, subset=None):
         if not self.index.is_monotonic:
             raise ValueError("asof requires a sorted index")
 
-        if isinstance(self, ABCSeries):
+        is_series = isinstance(self, ABCSeries)
+        if is_series:
             if subset is not None:
                 raise ValueError("subset is not valid for Series")
-            nulls = self.isnull()
         elif self.ndim > 2:
             raise NotImplementedError("asof is not implemented "
                                       "for {type}".format(type(self)))
@@ -3747,9 +3747,9 @@ def asof(self, where, subset=None):
                 subset = self.columns
             if not is_list_like(subset):
                 subset = [subset]
-            nulls = self[subset].isnull().any(1)
 
-        if not is_list_like(where):
+        is_list = is_list_like(where)
+        if not is_list:
             start = self.index[0]
             if isinstance(self.index, PeriodIndex):
                 where = Period(where, freq=self.index.freq).ordinal
@@ -3758,24 +3758,34 @@ def asof(self, where, subset=None):
             if where < start:
                 return np.nan
 
-            loc = self.index.searchsorted(where, side='right')
-            if loc > 0:
-                loc -= 1
-            while nulls[loc] and loc > 0:
-                loc -= 1
-            return self.iloc[loc]
+            # It's always much faster to use a *while* loop here for
+            # Series than pre-computing all the NAs. However a
+            # *while* loop is extremely expensive for DataFrame
+            # so we later pre-compute all the NAs and use the same
+            # code path whether *where* is a scalar or list.
+            # See PR: https://github.com/pandas-dev/pandas/pull/14476
+            if is_series:
+                loc = self.index.searchsorted(where, side='right')
+                if loc > 0:
+                    loc -= 1
+
+                values = self._values
+                while loc > 0 and isnull(values[loc]):
+                    loc -= 1
+                return values[loc]
 
         if not isinstance(where, Index):
-            where = Index(where)
+            where = Index(where) if is_list else Index([where])
 
+        nulls = self.isnull() if is_series else self[subset].isnull().any(1)
         locs = self.index.asof_locs(where, ~(nulls.values))
 
         # mask the missing
         missing = locs == -1
         data = self.take(locs, is_copy=False)
         data.index = where
         data.loc[missing] = np.nan
-        return data
+        return data if is_list else data.iloc[-1]
 
     # ----------------------------------------------------------------------
     # Action Methods