BUG: preserve datetime64 units during concat

Author: kushikks

pandas/core/dtypes/cast.py

@@ -447,6 +447,35 @@ def ensure_dtype_can_hold_na(dtype: DtypeObj) -> DtypeObj:
     type(np.nan): np.nan,
 }
 
+def maybe_align_dt64_units(arrs: list) -> list:
+    """
+    Align datetime64 arrays to the same unit *without* forcing conversion
+    to nanoseconds.
+    """
+    from pandas.core.dtypes.common import is_datetime64_dtype
+
+    if not any(is_datetime64_dtype(a.dtype) for a in arrs):
+        return arrs
+
+    def extract_unit(dtype):
+        name = dtype.name
+        if name.startswith("datetime64["):
+            return name[name.find("[")+1:-1]
+        return "ns"
+
+    units = [extract_unit(a.dtype) for a in arrs]
+
+    order = ["s", "ms", "us", "ns"]
+    target_unit = max(units, key=lambda u: order.index(u))
+
+    aligned = []
+    for arr in arrs:
+        if extract_unit(arr.dtype) != target_unit:
+            aligned.append(arr.astype(f"datetime64[{target_unit}]"))
+        else:
+            aligned.append(arr)
+
+    return aligned
 
 def maybe_promote(dtype: np.dtype, fill_value=np.nan):
     """

pandas/core/dtypes/concat.py

@@ -124,6 +124,10 @@ def concat_compat(
     else:
         to_concat_arrs = cast("Sequence[np.ndarray]", to_concat)
         result = np.concatenate(to_concat_arrs, axis=axis)
+        # PR FIX: preserve datetime64/timedelta64 units during concat
+        from pandas.core.dtypes.cast import maybe_align_dt64_units
+        result = maybe_align_dt64_units(result, to_concat_arrs)
+
 
         if not any_ea and "b" in kinds and result.dtype.kind in "iuf":
             # GH#39817 cast to object instead of casting bools to numeric

pandas/tests/reshape/concat/test_datetime_units.py

@@ -0,0 +1,42 @@
+import numpy as np
+import pandas as pd
+
+
+def test_concat_series_datetime64_different_units():
+    a = pd.Series(np.array(["2020-01-01T00:00:00"], dtype="datetime64[s]"))
+    b = pd.Series(np.array(["2020-01-02T00:00:00"], dtype="datetime64[ms]"))
+
+    result = pd.concat([a, b], ignore_index=True)
+
+    # dtype should still be datetime64 (M)
+    assert result.dtype.kind == "M"
+
+    assert result.iloc[0] == pd.Timestamp("2020-01-01")
+    assert result.iloc[1] == pd.Timestamp("2020-01-02")
+
+
+def test_concat_dataframe_datetime64_different_units():
+    df1 = pd.DataFrame({
+        "ts": np.array(["2020-01-01T00:00:00"], dtype="datetime64[s]")
+    })
+    df2 = pd.DataFrame({
+        "ts": np.array(["2020-01-02T00:00:00"], dtype="datetime64[ms]")
+    })
+
+    result = pd.concat([df1, df2], ignore_index=True)
+
+    assert result["ts"].dtype.kind == "M"
+    assert result["ts"].iloc[0] == pd.Timestamp("2020-01-01")
+    assert result["ts"].iloc[1] == pd.Timestamp("2020-01-02")
+
+
+def test_concat_datetime64_preserves_unit_order():
+    # Ensures that units are properly aligned, and no value shifts occur
+    a = pd.Series(np.array(["2020-01-01T00:00:00"], dtype="datetime64[us]"))
+    b = pd.Series(np.array(["2020-01-01T00:00:01"], dtype="datetime64[ns]"))
+
+    result = pd.concat([a, b], ignore_index=True)
+
+    assert result.dtype.kind == "M"
+    assert result.iloc[0] == pd.Timestamp("2020-01-01 00:00:00")
+    assert result.iloc[1] == pd.Timestamp("2020-01-01 00:00:01")