fix: table schema change error (#692)

* fix: table schema change error

* improve comments

Author: Linchin

pandas_gbq/gbq.py

@@ -1205,12 +1205,15 @@ def to_gbq(
         )
         table_connector.create(table_id, table_schema)
     else:
-        # Convert original schema (the schema that already exists) to pandas-gbq API format
-        original_schema = pandas_gbq.schema.to_pandas_gbq(table.schema)
-
-        # Update the local `table_schema` so mode (NULLABLE/REQUIRED)
-        # matches. See: https://github.com/pydata/pandas-gbq/issues/315
-        table_schema = pandas_gbq.schema.update_schema(table_schema, original_schema)
+        if if_exists == "append":
+            # Convert original schema (the schema that already exists) to pandas-gbq API format
+            original_schema = pandas_gbq.schema.to_pandas_gbq(table.schema)
+
+            # Update the local `table_schema` so mode (NULLABLE/REQUIRED)
+            # matches. See: https://github.com/pydata/pandas-gbq/issues/315
+            table_schema = pandas_gbq.schema.update_schema(
+                table_schema, original_schema
+            )
 
     if dataframe.empty:
         # Create the table (if needed), but don't try to run a load job with an

tests/system/test_gbq.py

@@ -788,6 +788,7 @@ def test_upload_data_if_table_exists_replace(self, project_id):
         test_size = 10
         df = make_mixed_dataframe_v2(test_size)
         df_different_schema = make_mixed_dataframe_v1()
+        schema_new = gbq.generate_bq_schema(df_different_schema)
 
         # Initialize table with sample data
         gbq.to_gbq(
@@ -798,7 +799,7 @@ def test_upload_data_if_table_exists_replace(self, project_id):
             credentials=self.credentials,
         )
 
-        # Test the if_exists parameter with the value 'replace'.
+        # When if_exists == 'replace', table schema should change too.
         gbq.to_gbq(
             df_different_schema,
             self.destination_table + test_id,
@@ -807,15 +808,16 @@ def test_upload_data_if_table_exists_replace(self, project_id):
             credentials=self.credentials,
         )
 
-        result = gbq.read_gbq(
-            "SELECT COUNT(*) AS num_rows FROM {0}".format(
-                self.destination_table + test_id
-            ),
+        df_new = gbq.read_gbq(
+            "SELECT * FROM {0}".format(self.destination_table + test_id),
             project_id=project_id,
             credentials=self.credentials,
             dialect="legacy",
         )
-        assert result["num_rows"][0] == 5
+
+        schema_returned = gbq.generate_bq_schema(df_new)
+        assert schema_new == schema_returned
+        assert df_new.shape[0] == 5
 
     def test_upload_data_if_table_exists_raises_value_error(self, project_id):
         test_id = "4"

tests/unit/test_gbq.py

@@ -650,6 +650,59 @@ def test_load_does_not_modify_schema_arg(mock_bigquery_client):
     assert original_schema == original_schema_cp
 
 
+def test_load_modifies_schema(mock_bigquery_client):
+    """Test of https://github.com/googleapis/python-bigquery-pandas/issues/670"""
+    from google.api_core.exceptions import NotFound
+
+    # Create table with new schema.
+    mock_bigquery_client.get_table.side_effect = NotFound("nope")
+    df = DataFrame(
+        {
+            "field1": ["a", "b"],
+            "field2": [1, 2],
+            "field3": [datetime.date(2019, 1, 1), datetime.date(2019, 5, 1)],
+        }
+    )
+    original_schema = [
+        {"name": "field1", "type": "STRING", "mode": "REQUIRED"},
+        {"name": "field2", "type": "INTEGER"},
+        {"name": "field3", "type": "DATE"},
+    ]
+    original_schema_cp = copy.deepcopy(original_schema)
+    gbq.to_gbq(
+        df,
+        "dataset.schematest",
+        project_id="my-project",
+        table_schema=original_schema,
+        if_exists="fail",
+    )
+    assert original_schema == original_schema_cp
+
+    # Test that when if_exists == "replace", the new table schema updates
+    # according to the local schema.
+    new_df = DataFrame(
+        {
+            "field1": ["a", "b"],
+            "field2": ["c", "d"],
+            "field3": [datetime.date(2019, 1, 1), datetime.date(2019, 5, 1)],
+        }
+    )
+    new_schema = [
+        {"name": "field1", "type": "STRING", "mode": "REQUIRED"},
+        {"name": "field2", "type": "STRING"},
+        {"name": "field3", "type": "DATE"},
+    ]
+    new_schema_cp = copy.deepcopy(new_schema)
+    gbq.to_gbq(
+        new_df,
+        "dataset.schematest",
+        project_id="my-project",
+        table_schema=new_schema,
+        if_exists="replace",
+    )
+    assert new_schema == new_schema_cp
+
+
 def test_read_gbq_passes_dtypes(mock_bigquery_client, mock_service_account_credentials):
     mock_service_account_credentials.project_id = "service_account_project_id"
     df = gbq.read_gbq(