diff --git a/bonobo_sqlalchemy/writers.py b/bonobo_sqlalchemy/writers.py
index 0734d41..537d53d 100644
--- a/bonobo_sqlalchemy/writers.py
+++ b/bonobo_sqlalchemy/writers.py
@@ -95,10 +95,7 @@ def commit(self, table, connection, buffer, force=False):
         if force or (buffer.qsize() >= self.buffer_size):
             with connection.begin():
                 while buffer.qsize() > 0:
-                    try:
-                        yield self.insert_or_update(table, connection, buffer.get())
-                    except Exception as exc:
-                        yield exc
+                    yield self.insert_or_update(table, connection, buffer.get())
 
     def insert_or_update(self, table, connection, row):
         """ Actual database load transformation logic, without the buffering / transaction logic. 
@@ -110,11 +107,12 @@ def insert_or_update(self, table, connection, row):
         # TODO XXX use actual database function instead of this stupid thing
         now = datetime.datetime.now()
 
+        target_row = row._asdict()
         column_names = table.columns.keys()
 
         # UpdatedAt field configured ? Let's set the value in source hash
         if self.updated_at_field in column_names:
-            row[self.updated_at_field] = now  # XXX not pure ...
+            target_row[self.updated_at_field] = now  # XXX not pure ...
 
         # Update logic
         if dbrow:
@@ -122,9 +120,9 @@ def insert_or_update(self, table, connection, row):
                 raise ProhibitedOperationError('UPDATE operations are not allowed by this transformation.')
 
             query = table.update().values(
-                **{col: row.get(col)
-                   for col in self.get_columns_for(column_names, row, dbrow)}
-            ).where(and_(*(getattr(table.c, col) == row.get(col) for col in self.discriminant)))
+                **{col: target_row.get(col)
+                   for col in self.get_columns_for(column_names, target_row, dbrow)}
+            ).where(and_(*(getattr(table.c, col) == target_row.get(col) for col in self.discriminant)))
 
         # INSERT
         else:
@@ -132,12 +130,12 @@ def insert_or_update(self, table, connection, row):
                 raise ProhibitedOperationError('INSERT operations are not allowed by this transformation.')
 
             if self.created_at_field in column_names:
-                row[self.created_at_field] = now  # XXX UNPURE
+                target_row[self.created_at_field] = now  # XXX UNPURE
             else:
-                if self.created_at_field in row:
-                    del row[self.created_at_field]  # UNPURE
+                if self.created_at_field in target_row:
+                    del target_row[self.created_at_field]  # UNPURE
 
-            query = table.insert().values(**{col: row.get(col) for col in self.get_columns_for(column_names, row)})
+            query = table.insert().values(**{col: target_row.get(col) for col in self.get_columns_for(column_names, target_row)})
 
         # Execute
         try:
@@ -175,7 +173,7 @@ def find(self, connection, table, row):
 
         return dict(row) if row else None
 
-    def get_columns_for(self, column_names, row, dbrow=None):
+    def get_columns_for(self, column_names, target_row, dbrow=None):
         """Retrieve list of table column names for which we have a value in given hash.
 
         """
@@ -184,12 +182,7 @@ def get_columns_for(self, column_names, row, dbrow=None):
         else:
             candidates = column_names
 
-        try:
-            fields = row._fields
-        except AttributeError as exc:
-            fields = list(row.keys())
-
-        return set(candidates).intersection(fields)
+        return set(candidates).intersection(target_row.keys())
 
     def add_fetch_columns(self, *columns, **aliased_columns):
         self.fetch_columns = {