fix: address review findings in Arrow path

- read_result_arrow(): drain response into BytesIO before releasing
  the connection, eliminating the fragile try/finally ordering
- Remove unused _FULL_TYPES_MAPPER dead code from arrow_convertor.py
- Remove stale docstring claiming primary_labels are always empty dicts
  (compute_primary_labels() now populates them)
- Add for_arrow_table() to DataFrameFactory class docstring
- convert_arrow_table_to_dataframe stub in __init__.py now raises
  ImportError with a helpful pyarrow install hint instead of being
  silently absent

Author: Martozar

gooddata-pandas/gooddata_pandas/__init__.py

@@ -6,7 +6,11 @@
 try:
     from gooddata_pandas.arrow_convertor import convert_arrow_table_to_dataframe
 except ImportError:
-    pass
+
+    def convert_arrow_table_to_dataframe(*args, **kwargs):  # type: ignore[misc]
+        raise ImportError("pyarrow is required for Arrow support. Install it with: pip install gooddata-pandas[arrow]")
+
+
 from gooddata_pandas.dataframe import DataFrameFactory
 from gooddata_pandas.good_pandas import GoodPandas
 from gooddata_pandas.result_convertor import LabelOverrides

gooddata-pandas/gooddata_pandas/arrow_convertor.py

@@ -23,23 +23,6 @@
     pa.large_string(): pandas.StringDtype("pyarrow"),
 }
 
-# Full nullable-type mapper: also maps integer and boolean Arrow types to their
-# pandas nullable equivalents.  NOT the default — nullable integer dtypes are
-# not backward compatible with the JSON path which produces float64 for all
-# numeric columns (with NaN for nulls).
-_FULL_TYPES_MAPPER: dict = {
-    **_ARROW_STRINGS_MAPPER,
-    pa.int8(): pandas.Int8Dtype(),
-    pa.int16(): pandas.Int16Dtype(),
-    pa.int32(): pandas.Int32Dtype(),
-    pa.int64(): pandas.Int64Dtype(),
-    pa.uint8(): pandas.UInt8Dtype(),
-    pa.uint16(): pandas.UInt16Dtype(),
-    pa.uint32(): pandas.UInt32Dtype(),
-    pa.uint64(): pandas.UInt64Dtype(),
-    pa.bool_(): pandas.BooleanDtype(),
-}
-
 
 def _label_ref_to_id_map(xtab_meta: dict) -> dict[str, str]:
     """Map 'l0', 'l1', ... to actual label local IDs."""

gooddata-pandas/gooddata_pandas/dataframe.py

@@ -70,6 +70,9 @@ class DataFrameFactory:
         - for_exec_def_arrow(self, exec_def: ExecutionDefinition,
             on_execution_submitted: Optional[Callable[[Execution], None]] = None)
             -> Tuple[pandas.DataFrame, DataFrameMetadata]:
+        - for_arrow_table(self, table: pa.Table,
+            execution_response: Optional[BareExecutionResponse] = None)
+            -> Tuple[pandas.DataFrame, DataFrameMetadata]:
         - for_exec_result_id(self, result_id: str, label_overrides: Optional[LabelOverrides] = None,
             result_cache_metadata: Optional[ResultCacheMetadata] = None,
             result_size_dimensions_limits: ResultSizeDimensions = (),
@@ -381,10 +384,6 @@ def for_exec_def_arrow(
         Returns the same ``(DataFrame, DataFrameMetadata)`` tuple as :meth:`for_exec_def` so that
         callers can switch between the two paths without changing their code.
 
-        ``DataFrameMetadata.primary_labels_from_index`` and
-        ``DataFrameMetadata.primary_labels_from_columns`` are always empty dicts because the
-        Arrow path does not support the ``use_primary_labels_in_attributes`` feature.
-
         Requires pyarrow to be installed (pip install gooddata-pandas[arrow]).
 
         Args:

gooddata-sdk/gooddata_sdk/compute/model/execution.py

@@ -387,16 +387,19 @@ def read_result_arrow(self) -> pyarrow.Table:
             raise ImportError(
                 "pyarrow is required to use read_result_arrow(). Install it with: pip install gooddata-sdk[arrow]"
             )
+        import io
+
         response = self._actions_api.retrieve_result_binary(
             workspace_id=self._workspace_id,
             result_id=self.result_id,
             _preload_content=False,
             **({"x_gdc_cancel_token": self.cancel_token} if self.cancel_token else {}),
         )
         try:
-            return _ipc.open_stream(response).read_all()
+            buf = io.BytesIO(response.read())
         finally:
             response.release_conn()
+        return _ipc.open_stream(buf).read_all()
 
     def cancel(self) -> None:
         """