[core] Reset next_task_reply_position when the actor restarts to avoid submitting the same task twice

[kevin85421]

Why are these changes needed?

Currently, we don't update the next_task_reply_position correctly. Hence, some tasks may belong to both out_of_order_completed_tasks and requests.

That is, when a actor restarts and calls ConnectActor, some tasks may be submitted by both ResendOutOfOrderCompletedTasks (out_of_order_completed_tasks) and SendPendingTasks (requests).

ray/src/ray/core_worker/transport/actor_task_submitter.cc

Lines 334 to 335 in 96a632d

	ResendOutOfOrderCompletedTasks(actor_id);
	SendPendingTasks(actor_id);

However, the Ray task status transitions to a different state during the first submission (ResendOutOfOrderCompletedTasks), so the second task submission (i.e., SendPendingTasks) fails on the following line.

ray/src/ray/core_worker/task_manager.cc

Line 1436 in 96a632d

RAY_CHECK(it->second.GetStatus() == rpc::TaskStatus::PENDING_NODE_ASSIGNMENT)

Example

Currently, we don't update the next_task_reply_position correctly. Hence, some tasks may belong to both out_of_order_completed_tasks and requests.

Add more details for the above statement. For example,

• Driver sends 100 tasks (seq_no: 0-99) to Actor A.
• 4 tasks finish and the driver receives the responses of seq_no (0 ~ 3). The driver calls SequentialActorSubmitQueue::MarkSeqnoCompleted 4 times to update next_task_reply_position to 4.
• Kill the worker Pod
• Reconstruct the objects by emplacing the first four tasks into the submission queue. Additionally, actor restarts also trigger the emplacement of the remaining 96 tasks. (seq_no: 100 ~ 199)
• 4 tasks finish and the driver receives the responses of seq_no (100 ~ 103). However, because next_task_reply_position is 4, these 4 tasks are put into out_of_order_completed_tasks.
• Kill the worker Pod
• Reconstruct the objects by emplacing the first four tasks into the submission queue. Additionally, actor restarts also trigger the emplacement of the remaining 96 tasks. (seq_no: 200 ~ 299). At the same time, the tasks (100 - 103) in the out_of_order_completed_tasks have already been updated to seq_no (200 - 203).
• Actor restarts and calls ConnectActor
  • tasks (200 - 203) are in both out_of_order_completed_tasks and requests.

Reproduction

Run the test without this PR

pytest -vs test_actor_lineage_reconstruction.py::test_mix_actor_restart_task_retry_and_lineage_reconstruction

Related issue number

Checks

• I've signed off every commit(by using the -s flag, i.e., git commit -s) in this PR.
• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
  • I've added any new APIs to the API Reference. For example, if I added a
    method in Tune, I've added it in doc/source/tune/api/ under the
    corresponding .rst file.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[kevin85421]

Some changes in this PR revert those introduced in #52249.