update comments

sky/backends/backend_utils.py

@@ -1839,9 +1839,8 @@ def run_ray_status_to_check_ray_cluster_healthy() -> bool:
     if (not node_statuses and handle.launched_resources.cloud.STATUS_VERSION >=
             clouds.StatusVersion.SKYPILOT):
         # Note: launched_at is set during sky launch, even on an existing
-        # cluster. This could cause extra checks if the cluster was already up
-        # before sky launch, but it will catch the case where the cluster was
-        # terminated on the cloud and restarted by sky launch.
+        # cluster. This will catch the case where the cluster was terminated on
+        # the cloud and restarted by sky launch.
         time_since_launch = time.time() - record['launched_at']
         if (record['status'] == status_lib.ClusterStatus.INIT and
                 time_since_launch < _LAUNCH_DOUBLE_CHECK_WINDOW):

sky/backends/cloud_vm_ray_backend.py

@@ -4133,7 +4133,14 @@ def post_teardown_cleanup(self,
         backend_utils.SSHConfigHelper.remove_cluster(handle.cluster_name)
 
         # Confirm that instances have actually transitioned state before
-        # updating the state database.
+        # updating the state database. We do this immediately before removing
+        # the state from the database, so that we can guarantee that this is
+        # always called before the state is removed. We considered running this
+        # check as part of provisioner.teardown_cluster or
+        # provision.terminate_instances, but it would open the door code paths
+        # that successfully call this function but do not first call
+        # teardown_cluster or terminate_instances. See
+        # https://github.com/skypilot-org/skypilot/pull/4443#discussion_r1872798032
         attempts = 0
         while True:
             logger.debug(f'instance statuses attempt {attempts + 1}')