fix(backends): ensure that the order_by expressions order is preserve…

…d when duplicates are introduced

ibis/backends/tests/test_window.py

@@ -1264,6 +1264,7 @@ def test_rank_followed_by_over_call_merge_frames(backend, alltypes, df):
     raises=sa.exc.ProgrammingError,
 )
 @pytest.mark.notimpl(["polars"], raises=com.OperationNotDefinedError)
+@pytest.mark.notyet(["flink"], raises=com.UnsupportedOperationError)
 @pytest.mark.broken(
     ["pyspark"], reason="pyspark requires CURRENT ROW", raises=PySparkAnalysisException
 )

ibis/expr/analysis.py

@@ -182,7 +182,14 @@ def merge_windows(_, default_frame):
     group_by = tuple(toolz.unique(_.frame.group_by + default_frame.group_by))
 
     order_by = {}
-    for sort_key in _.frame.order_by + default_frame.order_by:
+    # iterate in the order of the existing keys followed by the new keys
+    #
+    # this allows duplicates to be overridden with no effect on the original
+    # position
+    #
+    # see https://github.com/ibis-project/ibis/issues/7940 for how this
+    # originally manifested
+    for sort_key in default_frame.order_by + _.frame.order_by:
         order_by[sort_key.expr] = sort_key.ascending
     order_by = tuple(ops.SortKey(k, v) for k, v in order_by.items())
 

ibis/expr/types/generic.py

@@ -765,7 +765,6 @@ def bind(table):
                 )
             return expr
 
-        op = self.op()
         if isinstance(window, bl.WindowBuilder):
             if table := an.find_first_base_table(self.op()):
                 return bind(table)