Fix(optimizer): infinite recursion in the resolver (#7737)

Author: georgesittas

sqlglot/optimizer/resolver.py

@@ -157,9 +157,8 @@ def get_source_columns(self, name: str, only_visible: bool = False) -> Sequence[
             ):
                 columns = source_expr.named_selects
 
-                # in bigquery, unnest structs are automatically scoped as tables, so you can
-                # directly select a struct field in a query.
-                # this handles the case where the unnest is statically defined.
+                # in bigquery, unnest structs are automatically scoped as tables, so you can directly select
+                # a struct field in a query. This handles the case where the unnest is statically defined.
                 if self.dialect.UNNEST_COLUMN_ONLY and isinstance(source_expr, exp.Unnest):
                     if not source_expr.type or source_expr.type.is_type(exp.DType.UNKNOWN):
                         unnest_expr = seq_get(source_expr.expressions, 0)
@@ -178,7 +177,10 @@ def get_source_columns(self, name: str, only_visible: bool = False) -> Sequence[
                 ):
                     explode_col = source_expr.this.this
 
-                    if isinstance(explode_col, exp.Column) and source.parent:
+                    # If the column is unqualified at this point, it couldn't be resolved when
+                    # this scope's children were qualified; disambiguating it here would require
+                    # enumerating this very source's columns, i.e recurse without bound
+                    if isinstance(explode_col, exp.Column) and explode_col.table and source.parent:
                         col_type = self._get_unnest_column_type(explode_col, source.parent)
                         columns.extend(self._struct_field_names(col_type))
             elif isinstance(source, Scope) and isinstance(source.expression, exp.SetOperation):

tests/test_optimizer.py

@@ -750,6 +750,32 @@ def test_validate_columns(self):
         )
         self.assertEqual(expression.selects[0].type, exp.DataType.build("DOUBLE", dialect="spark"))
 
+        # An unqualified struct field is disambiguated through the lateral's extended columns
+        schema = {"my_table": {"items": "ARRAY<STRUCT<name STRING, age INT>>"}}
+        self.assertEqual(
+            optimizer.qualify.qualify(
+                parse_one(
+                    "SELECT name FROM my_table LATERAL VIEW EXPLODE(items) ci AS ci",
+                    read="spark",
+                ),
+                schema=schema,
+                dialect="spark",
+            ).sql(dialect="spark"),
+            "SELECT `ci`.`name` AS `name` FROM `my_table` AS `my_table` LATERAL VIEW EXPLODE(`my_table`.`items`) ci AS `ci`",
+        )
+
+        # Resolving an unqualified lateral column whose table is missing from the schema must
+        # raise instead of recursing infinitely
+        with self.assertRaisesRegex(OptimizeError, "Column 'ITEMS' could not be resolved"):
+            optimizer.qualify.qualify(
+                parse_one(
+                    "SELECT f.value AS v FROM my_db.raw.events, LATERAL FLATTEN(items) AS f",
+                    read="snowflake",
+                ),
+                schema={"my_db": {"other": {"some_view": {"v": "VARIANT"}}}},
+                dialect="snowflake",
+            )
+
     def test_qualify_columns__with_invisible(self):
         schema = MappingSchema(self.schema, {"x": {"a"}, "y": {"b"}, "z": {"b"}})
         self.check_file("qualify_columns__with_invisible", qualify_columns, schema=schema)