generic_loader/load_sas.py

@@ -1794,6 +1794,17 @@ def _prepare_for_copy(df: pd.DataFrame, columns: Dict[str, ColumnSpec]) -> pd.Da
                     series.replace({"": None}), errors="coerce"
                 )
                 out[name] = parsed.dt.date
+            elif pd.api.types.is_numeric_dtype(series):
+                # pyreadstat couldn't decode the SAS format (some
+                # ``DATEw.``/``YYMMDDw.`` variants and all custom formats slip
+                # through) so the column came back as float64: days since
+                # 1960-01-01, the SAS epoch. Without this branch the raw
+                # number would hit COPY and Postgres rejects it with
+                # ``invalid input syntax for type date``.
+                parsed = pd.to_datetime(
+                    series, unit="D", origin="1960-01-01", errors="coerce",
+                )
+                out[name] = parsed.dt.date
             else:
                 out[name] = series
         elif pg in ("TIMESTAMP", "TIMESTAMP WITHOUT TIME ZONE", "TIMESTAMP WITH TIME ZONE"):
@@ -1803,6 +1814,14 @@ def _prepare_for_copy(df: pd.DataFrame, columns: Dict[str, ColumnSpec]) -> pd.Da
                 out[name] = pd.to_datetime(
                     series.replace({"": None}), errors="coerce"
                 )
+            elif pd.api.types.is_numeric_dtype(series):
+                # Same story as the DATE branch above, but SAS datetimes are
+                # *seconds* since 1960-01-01 (fractional seconds for
+                # ``DATETIMEw.d``). Example caught in the wild:
+                # ``1915465463.615`` -> 2020-09-13 05:44:23.615.
+                out[name] = pd.to_datetime(
+                    series, unit="s", origin="1960-01-01", errors="coerce",
+                )
             else:
                 out[name] = series
         elif pg in ("TIME", "TIME WITHOUT TIME ZONE", "TIME WITH TIME ZONE"):