Tests and warnings for issue 399 (nan in table columns)

src/spatialdata/models/models.py

@@ -1061,6 +1061,21 @@ def _validate_table_annotation_metadata(self, data: AnnData) -> None:
         if len(set(expected_regions).symmetric_difference(set(found_regions))) > 0:
             raise ValueError(f"Regions in the AnnData object and `{attr[self.REGION_KEY_KEY]}` do not match.")
 
+        # Warning for object/string columns with NaN in region_key or instance_key
+        instance_key = attr[self.INSTANCE_KEY]
+        region_key = attr[self.REGION_KEY_KEY]
+        for key_name, key_value in [("region_key", region_key), ("instance_key", instance_key)]:
+            if key_value in data.obs:
+                col = data.obs[key_value]
+                col_dtype = col.dtype
+                if (col_dtype == "object" or pd.api.types.is_string_dtype(col_dtype)) and col.isna().any():
+                    logger.warning(
+                        f"The {key_name} column '{key_value}' is of {col_dtype} type and contains NaN values. "
+                        "After writing and reading with AnnData, NaN values may (depending on the AnnData version) "
+                        "be converted to strings. This may cause issues when matching instances across read/write "
+                        "cycles."
+                    )
+
     def validate(
         self,
         data: AnnData,

tests/io/test_readwrite.py

@@ -1065,3 +1065,45 @@ def test_read_sdata(tmp_path: Path, points: SpatialData) -> None:
     assert_spatial_data_objects_are_identical(sdata_from_path, sdata_from_str)
     assert_spatial_data_objects_are_identical(sdata_from_path, sdata_from_upath)
     assert_spatial_data_objects_are_identical(sdata_from_path, sdata_from_zarr_group)
+
+
+def test_sdata_with_nan_in_obs() -> None:
+    """Test writing SpatialData with mixed string/NaN values in obs works correctly.
+
+    Regression test for https://github.com/scverse/spatialdata/issues/399
+    Previously this raised TypeError: expected unicode string, found nan.
+    Now the write succeeds, though NaN values in object-dtype columns are
+    converted to the string "nan" after round-trip.
+    """
+    from spatialdata.models import TableModel
+
+    table = TableModel.parse(
+        AnnData(
+            obs=pd.DataFrame(
+                {
+                    "region": ["region1", "region2"],
+                    "instance": [0, 0],
+                    "column_only_region1": ["string", np.nan],
+                    "column_only_region2": [np.nan, 3],
+                }
+            )
+        ),
+        region_key="region",
+        instance_key="instance",
+        region=["region1", "region2"],
+    )
+    sdata = SpatialData(tables={"table": table})
+    assert sdata["table"].obs["column_only_region1"].iloc[1] is np.nan
+    assert np.isnan(sdata["table"].obs["column_only_region2"].iloc[0])
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        path = os.path.join(tmpdir, "data.zarr")
+        sdata.write(path)
+
+        sdata2 = SpatialData.read(path)
+        assert "column_only_region1" in sdata2["table"].obs.columns
+        assert sdata2["table"].obs["column_only_region1"].iloc[0] == "string"
+        assert sdata2["table"].obs["column_only_region2"].iloc[1] == 3
+        # After round-trip, NaN in object-dtype column becomes string "nan"
+        assert sdata2["table"].obs["column_only_region1"].iloc[1] == "nan"
+        assert np.isnan(sdata2["table"].obs["column_only_region2"].iloc[0])