perf(duckdb): speedup timestamp conversion by avoiding conversion to …

…object (#9556)

ibis/backends/duckdb/__init__.py

@@ -1393,7 +1393,7 @@ def execute(
                         # but calling `to_pylist()` will render it as None
                         col.null_count
                     )
-                    else col.to_pandas(timestamp_as_object=True)
+                    else col.to_pandas()
                 )
                 for name, col in zip(table.column_names, table.columns)
             }

ibis/backends/tests/test_asof_join.py

@@ -111,7 +111,8 @@ def test_asof_join(con, time_left, time_right, time_df1, time_df2, direction, op
     result = result.sort_values(["group", "time"]).reset_index(drop=True)
     expected = expected.sort_values(["group", "time"]).reset_index(drop=True)
 
-    tm.assert_frame_equal(result[expected.columns], expected)
+    # duckdb returns datetime64[us], pandas defaults to use datetime64[ns]
+    tm.assert_frame_equal(result[expected.columns], expected, check_dtype=False)
     with pytest.raises(AssertionError):
         tm.assert_series_equal(result["time"], result["time_right"])
 

ibis/formats/pandas.py

@@ -200,8 +200,10 @@ def convert_Boolean(cls, s, dtype, pandas_type):
 
     @classmethod
     def convert_Timestamp(cls, s, dtype, pandas_type):
-        if isinstance(dtype, pd.DatetimeTZDtype):
-            return s.dt.tz_convert(dtype.timezone)
+        if isinstance(pandas_type, pd.DatetimeTZDtype) and isinstance(
+            s.dtype, pd.DatetimeTZDtype
+        ):
+            return s if s.dtype == pandas_type else s.dt.tz_convert(dtype.timezone)
         elif pdt.is_datetime64_dtype(s.dtype):
             return s.dt.tz_localize(dtype.timezone)
         else:

ibis/tests/benchmarks/test_benchmarks.py

@@ -8,11 +8,13 @@
 import os
 import random
 import string
+from datetime import datetime
 from operator import attrgetter, itemgetter
 
 import numpy as np
 import pandas as pd
 import pytest
+import pytz
 from packaging.version import parse as vparse
 from pytest import param
 
@@ -914,3 +916,15 @@ def test_wide_drop_compile(benchmark, wide_table, cols_to_drop):
     benchmark(
         lambda expr: ibis.to_sql(expr, dialect="duckdb"), wide_table.drop(*cols_to_drop)
     )
+
+
+def test_duckdb_timestamp_conversion(benchmark):
+    pytest.importorskip("duckdb")
+
+    start = datetime(2000, 1, 1, tzinfo=pytz.UTC)
+    stop = datetime(2000, 2, 1, tzinfo=pytz.UTC)
+    expr = ibis.range(start, stop, ibis.interval(seconds=1)).unnest()
+
+    con = ibis.duckdb.connect()
+    series = benchmark(con.execute, expr)
+    assert series.size == (stop - start).total_seconds()