fix: support melting empty DataFrames without crashing (#2509)

Allows alignment melts over zero-row offset layouts
Fixes #<452681068> 🦕

Author: shuoweil

bigframes/core/blocks.py

@@ -1822,9 +1822,9 @@ def melt(
  Arguments correspond to pandas.melt arguments.
  """
 # TODO: Implement col_level and ignore_index
-value_labels: pd.Index=pd.Index(
- [self.col_id_to_label[col_id]forcol_idinvalue_vars]
-)
+value_labels: pd.Index=self.column_labels[
+ [self.value_columns.index(col_id)forcol_idinvalue_vars]
+]
 id_labels= [self.col_id_to_label[col_id] forcol_idinid_vars]
 
 unpivot_expr, (var_col_ids, unpivot_out, passthrough_cols) =unpivot(
@@ -3417,6 +3417,7 @@ def unpivot(
 joined_array, (labels_mapping, column_mapping) =labels_array.relational_join(
 array_value, type="cross"
  )
+
 new_passthrough_cols= [column_mapping[col] forcolinpassthrough_columns]
 # Last column is offsets
 index_col_ids= [labels_mapping[col] forcolinlabels_array.column_ids[:-1]]
@@ -3426,20 +3427,24 @@ def unpivot(
 unpivot_exprs: List[ex.Expression] = []
 # Supports producing multiple stacked ouput columns for stacking only part of hierarchical index
 forinput_idsinunpivot_columns:
-# row explode offset used to choose the input column
-# we use offset instead of label as labels are not necessarily unique
-cases=itertools.chain(
-*(
- (
-ops.eq_op.as_expr(explode_offsets_id, ex.const(i)),
-ex.deref(column_mapping[id_or_null])
-if (id_or_nullisnotNone)
-elseex.const(None),
+col_expr: ex.Expression
+ifnotinput_ids:
+col_expr=ex.const(None, dtype=bigframes.dtypes.INT_DTYPE)
+else:
+# row explode offset used to choose the input column
+# we use offset instead of label as labels are not necessarily unique
+cases=itertools.chain(
+*(
+ (
+ops.eq_op.as_expr(explode_offsets_id, ex.const(i)),
+ex.deref(column_mapping[id_or_null])
+if (id_or_nullisnotNone)
+elseex.const(None),
+ )
+fori, id_or_nullinenumerate(input_ids)
  )
-fori, id_or_nullinenumerate(input_ids)
  )
- )
-col_expr=ops.case_when_op.as_expr(*cases)
+col_expr=ops.case_when_op.as_expr(*cases)
 unpivot_exprs.append(col_expr)
 
 joined_array, unpivot_col_ids=joined_array.compute_values(unpivot_exprs)
@@ -3457,19 +3462,43 @@ def _pd_index_to_array_value(
  Create an ArrayValue from a list of label tuples.
  The last column will be row offsets.
  """
+id_gen=bigframes.core.identifiers.standard_id_strings()
+col_ids= [next(id_gen) for_inrange(index.nlevels)]
+offset_id=next(id_gen)
+
 rows= []
 labels_as_tuples=utils.index_as_tuples(index)
 forrow_offsetinrange(len(index)):
-id_gen=bigframes.core.identifiers.standard_id_strings()
 row_label=labels_as_tuples[row_offset]
 row_label= (row_label,) ifnotisinstance(row_label, tuple) elserow_label
 row= {}
-forlabel_part, idinzip(row_label, id_gen):
-row[id] =label_partifpd.notnull(label_part) elseNone
-row[next(id_gen)] =row_offset
+forlabel_part, col_idinzip(row_label, col_ids):
+row[col_id] =label_partifpd.notnull(label_part) elseNone
+row[offset_id] =row_offset
 rows.append(row)
 
-returncore.ArrayValue.from_pyarrow(pa.Table.from_pylist(rows), session=session)
+ifnotrows:
+dtypes_list=getattr(index, "dtypes", None)
+ifdtypes_listisNone:
+dtypes_list= (
+ [index.dtype] ifhasattr(index, "dtype") else [pd.Float64Dtype()]
+ )
+
+fields= []
+forcol_id, dtypeinzip(col_ids, dtypes_list):
+try:
+pa_type=bigframes.dtypes.bigframes_dtype_to_arrow_dtype(dtype)
+exceptException:
+pa_type=pa.string()
+fields.append(pa.field(col_id, pa_type))
+fields.append(pa.field(offset_id, pa.int64()))
+schema=pa.schema(fields)
+pt=pa.Table.from_pylist([], schema=schema)
+else:
+pt=pa.Table.from_pylist(rows)
+pt=pt.rename_columns([*col_ids, offset_id])
+
+returncore.ArrayValue.from_pyarrow(pt, session=session)
 
 
 def_resolve_index_col(

tests/system/small/test_dataframe.py

@@ -5902,6 +5902,19 @@ def test_to_gbq_table_labels(scalars_df_index):
 asserttable.labels["test"] =="labels"
 
 
+deftest_to_gbq_obj_ref_persists(session):
+# Test that saving and loading an Object Reference retains its dtype
+bdf=session.from_glob_path(
+"gs://cloud-samples-data/vision/ocr/*.jpg", name="uris"
+ ).head(1)
+
+destination_table="bigframes-dev.bigframes_tests_sys.test_obj_ref_persistence"
+bdf.to_gbq(destination_table, if_exists="replace")
+
+loaded_df=session.read_gbq(destination_table)
+assertloaded_df["uris"].dtype==dtypes.OBJ_REF_DTYPE
+
+
 @pytest.mark.parametrize(
  ("col_names", "ignore_index"),
  [

tests/system/small/test_multiindex.py

@@ -1490,3 +1490,34 @@ def test_multiindex_eq_const(scalars_df_index, scalars_pandas_df_index):
 bigframes.testing.utils.assert_index_equal(
 pandas.Index(pd_result, dtype="boolean"), bf_result.to_pandas()
  )
+
+
+deftest_count_empty_multiindex_columns(session):
+df=pandas.DataFrame(
+ [], index=[1, 2], columns=pandas.MultiIndex.from_tuples([], names=["a", "b"])
+ )
+bdf=session.read_pandas(df)
+
+# count() operation unpivots columns, triggering the empty MultiIndex bug internally
+count_df=bdf.count()
+
+# The local fix ensures that empty unpivoted columns generate properly typed NULLs
+# rather than failing syntax validation downstream in BigQuery.
+# We compile to `.sql` to verify it succeeds locally without evaluating on BigQuery natively.
+_=count_df.to_frame().sql
+
+# Assert structural layout is correct
+assertcount_df.index.nlevels==2
+assertlist(count_df.index.names) == ["a", "b"]
+
+
+deftest_dataframe_melt_multiindex(session):
+# Tests that `melt` operations via count do not cause MultiIndex drops in Arrow
+df=pandas.DataFrame({"A": [1], "B": ["string"], "C": [3]})
+df.columns=pandas.MultiIndex.from_tuples(
+ [("Group1", "A"), ("Group2", "B"), ("Group1", "C")]
+ )
+bdf=session.read_pandas(df)
+
+count_df=bdf.count().to_pandas()
+assertcount_df.shape[0] ==3