fix: stop DIALECT_PATTERN backtracking over runs of block comments

sqlmesh/core/dialect.py

@@ -1124,7 +1124,14 @@ def text_diff(
     return "\n".join(unified_diff(a_sql, b_sql))
 
 
-WS_OR_COMMENT = r"(?:\s|--[^\n]*\n|/\*.*?\*/)"
+# Each '*/' closes exactly one '/*', so a run of comments matches in exactly one way. A lazy
+# '/\*.*?\*/' under DOTALL can stretch across several comments, and a failed search then
+# backtracks through every grouping of them. SQLGlot reads block comments as nested, so this
+# matches nesting up to three levels deep; a regex cannot match arbitrary depth.
+BLOCK_COMMENT = r"/\*(?:[^*]|\*(?!/))*\*/"
+for _ in range(2):
+    BLOCK_COMMENT = rf"/\*(?:[^*/]|\*(?!/)|/(?!\*)|{BLOCK_COMMENT})*\*/"
+WS_OR_COMMENT = rf"(?:\s|--[^\n]*\n|{BLOCK_COMMENT})"
 HEADER = r"\b(?:model|audit)\b(?=\s*\()"
 KEY_BOUNDARY = r"(?:\(|,)"  # key is preceded by either '(' or ','
 DIALECT_VALUE = r"['\"]?(?P<dialect>[a-z][a-z0-9]*)['\"]?"

tests/core/test_model.py

@@ -3151,6 +3151,26 @@ def assert_match(test_sql: str, expected_value: t.Optional[str] = "duckdb"):
         """
     )
 
+    # a run of block comments before the dialect key
+    comments = "\n".join(f"/* note {i} */" for i in range(40))
+    assert_match(make_test_sql(f"tag my_tag, {comments}\ndialect duckdb"))
+
+    # a run of block comments and no dialect key: the search fails without backtracking
+    # through every grouping of the comments
+    test_sql = make_test_sql(f"tag my_tag, {comments}\ncron '@daily'")
+    matches = list(d.DIALECT_PATTERN.finditer(test_sql))
+    assert not matches
+
+    # nested block comments before the dialect key
+    assert_match(
+        make_test_sql(
+            """
+            tag my_tag, /* outer /* inner /* innermost */ */ still outer */
+            dialect duckdb
+            """
+        )
+    )
+
     # full parse
     sql = """
     MODEL (