COMPAT: prepare for pandas 3.0 string dtype

.github/workflows/tests-conda.yml

@@ -44,6 +44,7 @@ jobs:
           - os: "ubuntu-latest"
             python: "3.11"
             env: "nightly-deps"
+            pandas_future_infer_string: "1"
 
     steps:
       - name: Checkout repo
@@ -68,5 +69,7 @@ jobs:
         run: pip install -e .
 
       - name: Test
+        env:
+          PANDAS_FUTURE_INFER_STRING: ${{ matrix.pandas_future_infer_string || '0' }}
         run: |
-          pytest -v --color=yes -r s pyogrio/tests
+          pytest -v -s --color=yes -r s pyogrio/tests

pyogrio/_compat.py

@@ -33,16 +33,21 @@
 HAS_ARROW_WRITE_API = __gdal_version__ >= (3, 8, 0)
 HAS_PYARROW = pyarrow is not None
 HAS_PYPROJ = pyproj is not None
+PYARROW_GE_19 = pyarrow is not None and Version(pyarrow.__version__) >= Version(
+    "19.0.0"
+)
 
 HAS_GEOPANDAS = geopandas is not None
 
 PANDAS_GE_15 = pandas is not None and Version(pandas.__version__) >= Version("1.5.0")
 PANDAS_GE_20 = pandas is not None and Version(pandas.__version__) >= Version("2.0.0")
 PANDAS_GE_22 = pandas is not None and Version(pandas.__version__) >= Version("2.2.0")
+PANDAS_GE_30 = pandas is not None and Version(pandas.__version__) >= Version("3.0.0dev")
 
 GDAL_GE_352 = __gdal_version__ >= (3, 5, 2)
 GDAL_GE_37 = __gdal_version__ >= (3, 7, 0)
 GDAL_GE_38 = __gdal_version__ >= (3, 8, 0)
+GDAL_GE_311 = __gdal_version__ >= (3, 11, 0)
 
 HAS_GDAL_GEOS = __gdal_geos_version__ is not None
 

pyogrio/geopandas.py

@@ -5,7 +5,14 @@
 
 import numpy as np
 
-from pyogrio._compat import HAS_GEOPANDAS, PANDAS_GE_15, PANDAS_GE_20, PANDAS_GE_22
+from pyogrio._compat import (
+    HAS_GEOPANDAS,
+    PANDAS_GE_15,
+    PANDAS_GE_20,
+    PANDAS_GE_22,
+    PANDAS_GE_30,
+    PYARROW_GE_19,
+)
 from pyogrio.errors import DataSourceError
 from pyogrio.raw import (
     DRIVERS_NO_MIXED_DIMENSIONS,
@@ -52,13 +59,13 @@ def _try_parse_datetime(ser):
         except Exception:
             res = ser
     # if object dtype, try parse as utc instead
-    if res.dtype == "object":
+    if res.dtype in ("object", "string"):
         try:
             res = pd.to_datetime(ser, utc=True, **datetime_kwargs)
         except Exception:
             pass
 
-    if res.dtype != "object":
+    if res.dtype.kind == "M":
         # GDAL only supports ms precision, convert outputs to match.
         # Pandas 2.0 supports datetime[ms] directly, prior versions only support [ns],
         # Instead, round the values to [ms] precision.
@@ -285,11 +292,31 @@ def read_dataframe(
     )
 
     if use_arrow:
+        import pyarrow as pa
+
         meta, table = result
 
         # split_blocks and self_destruct decrease memory usage, but have as side effect
         # that accessing table afterwards causes crash, so del table to avoid.
         kwargs = {"self_destruct": True}
+        if PANDAS_GE_30:
+            # starting with pyarrow 19.0, pyarrow will correctly handle this themselves,
+            # so only use types_mapper as workaround for older versions
+            if not PYARROW_GE_19:
+                kwargs["types_mapper"] = {
+                    pa.string(): pd.StringDtype(na_value=np.nan),
+                    pa.large_string(): pd.StringDtype(na_value=np.nan),
+                    pa.json_(): pd.StringDtype(na_value=np.nan),
+                }.get
+            # TODO enable the below block when upstream issue to accept extension types
+            # is fixed
+            # else:
+            #     # for newer pyarrow, still include mapping for json
+            #     # GDAL 3.11 started to emit this extension type, but pyarrow does not
+            #     # yet support it properly in the conversion to pandas
+            #     kwargs["types_mapper"] = {
+            #         pa.json_(): pd.StringDtype(na_value=np.nan),
+            #     }.get
         if arrow_to_pandas_kwargs is not None:
             kwargs.update(arrow_to_pandas_kwargs)
         df = table.to_pandas(**kwargs)

pyogrio/tests/test_geopandas_io.py

@@ -18,10 +18,12 @@
 )
 from pyogrio._compat import (
     GDAL_GE_37,
+    GDAL_GE_311,
     GDAL_GE_352,
     HAS_ARROW_WRITE_API,
     HAS_PYPROJ,
     PANDAS_GE_15,
+    PANDAS_GE_30,
     SHAPELY_GE_21,
 )
 from pyogrio.errors import DataLayerError, DataSourceError, FeatureError, GeometryError
@@ -256,13 +258,20 @@ def test_read_layer(tmp_path, use_arrow):
 
     # create a multilayer GPKG
     expected1 = gp.GeoDataFrame(geometry=[Point(0, 0)], crs="EPSG:4326")
+    if use_arrow:
+        # TODO this needs to be fixed on the geopandas side (to ensure the
+        # GeoDataFrame() constructor does this), when use_arrow we already
+        # get columns Index with string dtype
+        expected1.columns = expected1.columns.astype("str")
     write_dataframe(
         expected1,
         filename,
         layer="layer1",
     )
 
     expected2 = gp.GeoDataFrame(geometry=[Point(1, 1)], crs="EPSG:4326")
+    if use_arrow:
+        expected2.columns = expected2.columns.astype("str")
     write_dataframe(expected2, filename, layer="layer2", append=True)
 
     assert np.array_equal(
@@ -385,7 +394,7 @@ def test_read_null_values(tmp_path, use_arrow):
     df = read_dataframe(filename, use_arrow=use_arrow, read_geometry=False)
 
     # make sure that Null values are preserved
-    assert np.array_equal(df.col.values, expected.col.values)
+    assert df["col"].isna().all()
 
 
 def test_read_fid_as_index(naturalearth_lowres_all_ext, use_arrow):
@@ -699,6 +708,13 @@ def test_read_skip_features(naturalearth_lowres_all_ext, use_arrow, skip_feature
     # In .geojsonl the vertices are reordered, so normalize
     is_jsons = ext == ".geojsonl"
 
+    if skip_features == 200 and not use_arrow:
+        # result is an empty dataframe, so no proper dtype inference happens
+        # for the numpy object dtype arrays
+        df[["continent", "name", "iso_a3"]] = df[
+            ["continent", "name", "iso_a3"]
+        ].astype("str")
+
     assert_geodataframe_equal(
         df,
         expected,
@@ -1180,6 +1196,10 @@ def test_write_empty_dataframe(tmp_path, ext, columns, dtype, use_arrow):
     # For older pandas versions, the index is created as Object dtype but read as
     # RangeIndex, so don't check the index dtype in that case.
     check_index_type = True if PANDAS_GE_20 else False
+    # with pandas 3+ and reading through arrow, we preserve the string dtype
+    # (no proper dtype inference happens for the empty numpy object dtype arrays)
+    if use_arrow and dtype is object:
+        expected["col_object"] = expected["col_object"].astype("str")
     assert_geodataframe_equal(df, expected, check_index_type=check_index_type)
 
 
@@ -1214,7 +1234,11 @@ def test_write_None_string_column(tmp_path, use_arrow):
     assert filename.exists()
 
     result_gdf = read_dataframe(filename, use_arrow=use_arrow)
-    assert result_gdf.object_col.dtype == object
+    assert (
+        result_gdf.object_col.dtype == "str" if PANDAS_GE_30 and use_arrow else object
+    )
+    if PANDAS_GE_30 and use_arrow:
+        gdf["object_col"] = gdf["object_col"].astype("str")
     assert_geodataframe_equal(result_gdf, gdf)
 
 
@@ -1658,11 +1682,12 @@ def test_write_read_mixed_column_values(tmp_path):
     write_dataframe(test_gdf, output_path)
     output_gdf = read_dataframe(output_path)
     assert len(test_gdf) == len(output_gdf)
-    for idx, value in enumerate(mixed_values):
-        if value in (None, np.nan):
-            assert output_gdf["mixed"][idx] is None
-        else:
-            assert output_gdf["mixed"][idx] == str(value)
+    # mixed values as object dtype are currently written as strings
+    expected = pd.Series(
+        [str(value) if value not in (None, np.nan) else None for value in mixed_values],
+        name="mixed",
+    )
+    assert_series_equal(output_gdf["mixed"], expected)
 
 
 @requires_arrow_write_api
@@ -1695,8 +1720,8 @@ def test_write_read_null(tmp_path, use_arrow):
     assert pd.isna(result_gdf["float64"][1])
     assert pd.isna(result_gdf["float64"][2])
     assert result_gdf["object_str"][0] == "test"
-    assert result_gdf["object_str"][1] is None
-    assert result_gdf["object_str"][2] is None
+    assert pd.isna(result_gdf["object_str"][1])
+    assert pd.isna(result_gdf["object_str"][2])
 
 
 @pytest.mark.requires_arrow_write_api
@@ -1927,6 +1952,8 @@ def test_read_dataset_kwargs(nested_geojson_file, use_arrow):
         geometry=[shapely.Point(0, 0)],
         crs="EPSG:4326",
     )
+    if GDAL_GE_311 and use_arrow:
+        expected["intermediate_level"] = expected["intermediate_level"].astype(object)
 
     assert_geodataframe_equal(df, expected)
 
@@ -1972,7 +1999,7 @@ def test_write_nullable_dtypes(tmp_path, use_arrow):
     expected["col2"] = expected["col2"].astype("float64")
     expected["col3"] = expected["col3"].astype("float32")
     expected["col4"] = expected["col4"].astype("float64")
-    expected["col5"] = expected["col5"].astype(object)
+    expected["col5"] = expected["col5"].astype("str")
     expected.loc[1, "col5"] = None  # pandas converts to pd.NA on line above
     assert_geodataframe_equal(output_gdf, expected)