fix(db): memoise InvalidNameError GUC rejections like UndefinedObjectError

hindsight-api-slim/hindsight_api/engine/db/postgresql.py

@@ -72,15 +72,22 @@ async def apply_session_settings(conn: asyncpg.Connection, settings: list[tuple[
     for name, value in settings:
         try:
             await conn.execute("SELECT set_config($1, $2, false)", name, value)
-        except asyncpg.exceptions.UndefinedObjectError:
+        except (asyncpg.exceptions.UndefinedObjectError, asyncpg.exceptions.InvalidNameError):
             # The server does not define this GUC — an extension we tune for is absent
-            # or predates it (hnsw.iterative_scan needs pgvector 0.8+, and pgvector
-            # reserves the "hnsw." prefix, so an older one rejects it rather than
-            # accepting a placeholder). Remember it: otherwise every acquire from here
-            # on re-pays a failed batch plus one statement per setting, which behind a
-            # transaction-mode pooler is a server-side transaction each — the burn
-            # #3499 removed. Narrow to UndefinedObjectError so a transient failure
-            # does not disable a setting the server does support.
+            # or predates it (hnsw.iterative_scan needs pgvector 0.8+). Which error
+            # that raises depends on the extension: a plain unknown name is 42704
+            # UndefinedObjectError ("unrecognized configuration parameter"), while a
+            # name under a prefix the loaded extension has reserved is 42602
+            # InvalidNameError ("invalid configuration parameter name") — pgvector
+            # 0.6.0 answers the latter for hnsw.iterative_scan. Both verdicts are
+            # permanent for this server, so remember either: otherwise every acquire
+            # from here on re-pays a failed batch plus one statement per setting,
+            # which behind a transaction-mode pooler is a server-side transaction
+            # each — the burn #3499 removed — and, worse, setting_rejected_by_server
+            # keeps answering False, so retain's link probing SET LOCALs the GUC
+            # inside its own transaction and aborts the whole link computation.
+            # Still narrow (no bare PostgresError) so a transient failure does not
+            # disable a setting the server does support.
             logger.info("Server does not know %s — not sending it again on this process", name)
             _unsupported_settings.add(name)
         except asyncpg.exceptions.PostgresError:

hindsight-api-slim/tests/test_db_abstraction.py

@@ -645,17 +645,23 @@ async def cb(conn):
 class _RecordingConnection:
     """Captures every statement, optionally failing the first (batched) one."""
 
-    def __init__(self, fail_batched: bool = False, reject: str | None = None) -> None:
+    def __init__(
+        self,
+        fail_batched: bool = False,
+        reject: str | None = None,
+        reject_error: type[Exception] = asyncpg.exceptions.UndefinedObjectError,
+    ) -> None:
         self.calls: list[tuple[str, tuple]] = []
         self._fail_batched = fail_batched
         self._reject = reject
+        self._reject_error = reject_error
 
     async def execute(self, query: str, *args) -> None:
         self.calls.append((query, args))
         if self._fail_batched and len(self.calls) == 1:
-            raise asyncpg.exceptions.UndefinedObjectError("unrecognized configuration parameter")
+            raise self._reject_error("unrecognized configuration parameter")
         if self._reject is not None and self._reject in args:
-            raise asyncpg.exceptions.UndefinedObjectError("unrecognized configuration parameter")
+            raise self._reject_error("unrecognized configuration parameter")
 
 
 class TestApplySessionSettings:
@@ -727,6 +733,36 @@ async def test_a_setting_the_server_rejects_is_not_sent_again(self):
         assert "pg_trgm.similarity_threshold" not in args
         assert args == ("hnsw.ef_search", "200", "statement_timeout", "600s")
 
+    @pytest.mark.asyncio
+    async def test_invalid_name_rejection_is_remembered_like_undefined_object(self):
+        """A reserved-prefix GUC rejection (InvalidNameError) must also be permanent.
+
+        The docstring above assumes an old server answers "unrecognized configuration
+        parameter" (UndefinedObjectError, 42704). PG16 + pgvector 0.6.0 answers
+        `invalid configuration parameter name "hnsw.iterative_scan"` instead —
+        InvalidNameError (42602), because the loaded extension reserves the "hnsw."
+        prefix but predates the GUC. If only UndefinedObjectError is remembered,
+        the name never reaches the unsupported-set, ``setting_rejected_by_server``
+        keeps answering False, and retain's link probing sends the GUC via SET LOCAL
+        inside its own transaction — aborting the whole link computation on every
+        retain (observed in production on PG16 + pgvector 0.6.0, 2026-08-26).
+        """
+        conn = _RecordingConnection(
+            fail_batched=True,
+            reject="hnsw.ef_search",
+            reject_error=asyncpg.exceptions.InvalidNameError,
+        )
+        await apply_session_settings(conn, self._SETTINGS)
+
+        assert pg_backend.setting_rejected_by_server("hnsw.ef_search")
+
+        # Next acquire must not re-send the rejected name.
+        conn = _RecordingConnection()
+        await apply_session_settings(conn, self._SETTINGS)
+        assert len(conn.calls) == 1
+        _, args = conn.calls[0]
+        assert "hnsw.ef_search" not in args
+
     @pytest.mark.asyncio
     async def test_a_transient_failure_does_not_disable_a_setting(self):
         """Only "unrecognized configuration parameter" is permanent; anything else retries."""