Fix actor creation hang due to race in SWAP queue (#6280)

src/ray/core_worker/core_worker.cc

@@ -709,9 +709,9 @@ bool CoreWorker::AddActorHandle(std::unique_ptr<ActorHandle> actor_handle) {
           it->second->Reset();
         }
       } else if (actor_data.state() == gcs::ActorTableData::DEAD) {
-        RAY_CHECK_OK(gcs_client_->Actors().AsyncUnsubscribe(actor_id, nullptr));
         // We cannot erase the actor handle here because clients can still
-        // submit tasks to dead actors.
+        // submit tasks to dead actors. This also means we defer unsubscription,
+        // otherwise we crash when bulk unsubscribing all actor handles.
       }
 
       direct_actor_submitter_->HandleActorUpdate(actor_id, actor_data);

src/ray/raylet/node_manager.cc

@@ -2632,21 +2632,9 @@ void NodeManager::ForwardTask(
         task_entry.second.TaskData().GetTaskExecutionSpec().GetMessage());
   }
 
-  // Move the FORWARDING task to the SWAP queue so that we remember that we
-  // have it queued locally. Once the ForwardTaskRequest has been sent, the
-  // task will get re-queued, depending on whether the message succeeded or
-  // not.
-  local_queues_.QueueTasks({task}, TaskState::SWAP);
-  client->ForwardTask(request, [this, on_error, task_id, node_id](
+  client->ForwardTask(request, [this, on_error, task, task_id, node_id](
                                    Status status, const rpc::ForwardTaskReply &reply) {
     // Remove the FORWARDING task from the SWAP queue.
-    Task task;
-    TaskState state;
-    if (!local_queues_.RemoveTask(task_id, &task, &state)) {
-      return;
-    }
-    RAY_CHECK(state == TaskState::SWAP);
-
     if (status.ok()) {
       const auto &spec = task.GetTaskSpecification();
       // Mark as forwarded so that the task and its lineage are not