[core] Always create a default executor #51058

kevin85421 · 2025-03-04T08:39:29Z

Why are these changes needed?

Currently, for non-threaded actors, Ray tasks are executed on the main thread. At the same time, the main thread is also responsible for control plane logic such as scheduling and receiving RPCs. In other words, for non-threaded actors, the control plane and computation logic are coupled. However, for threaded actors, they are decoupled.

This PR always creates a default executor to decouple the control plane from the computation logic.

Currently, some Ray features (See all the PRs where I fixed the CI failures to unblock this PR.) rely on coupling the control plane and computation logic, which doesn't work for threaded actors. With this PR, we no longer need to worry about features working exclusively on non-threaded actors instead of on threaded actors because Ray tasks are always running in an executor.

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 · 2025-03-04T08:40:29Z

src/ray/core_worker/transport/task_receiver.cc

@@ -169,19 +169,17 @@ void TaskReceiver::HandleTask(const rpc::PushTaskRequest &request,
      }

      if (task_spec.IsActorCreationTask()) {
-        /// The default max concurrency for creating PoolManager should


If an actor is async, pool_manager_ will never be used.

ray/src/ray/core_worker/transport/actor_scheduling_queue.cc

Lines 175 to 194 in 584d826

if (is_asyncio_) {

// Process async actor task.

auto fiber = fiber_state_manager_->GetExecutor(request.ConcurrencyGroupName(),

request.FunctionDescriptor());

fiber->EnqueueFiber([this, request, task_id]() mutable {

AcceptRequestOrRejectIfCanceled(task_id, request);

});

} else {

// Process actor tasks.

RAY_CHECK(pool_manager_ != nullptr);

auto pool = pool_manager_->GetExecutor(request.ConcurrencyGroupName(),

request.FunctionDescriptor());

if (pool == nullptr) {

AcceptRequestOrRejectIfCanceled(task_id, request);

} else {

pool->Post([this, request, task_id]() mutable {

AcceptRequestOrRejectIfCanceled(task_id, request);

});

}

}

kevin85421 · 2025-03-04T21:59:44Z

There are some CI failures. Separate a part of this PR into #51069 to make this PR smaller to review and debug.

kevin85421 · 2025-03-06T09:55:07Z

note for myself:

setproctitle doesn't work as our expectation for a child actor and the task is running on non-main thread.

cmdline is udpated as expected
/proc/pid/comm is not updated --> not our expectations.
/proc/pid/task/tid/comm is updated (we didn't know before): PR_SET_NAME

"title" has multiple definitions in setproctitle depends on the dev envs. However, we only use one of them which is not correct. --> address this assumption first

The process title is usually visible in files such as /proc/PID/cmdline, /proc/PID/status, /proc/PID/comm, depending on the operating system and kernel version. These information are used by user-space tools such as ps and top.

Signed-off-by: kaihsun <kaihsun@anyscale.com>

…tmux-0-ray6

kevin85421 · 2025-03-09T19:00:09Z

https://github.com/python/cpython/blob/475f933ed8b1c9546f1b5497a2241140c7065b5f/Python/pystate.c#L2864

kevin85421 · 2025-03-09T21:49:41Z

[Done] This PR is currently blocked by #51203

…tmux-0-ray6

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 · 2025-03-13T01:27:36Z

python/ray/tests/test_task_events.py

@@ -333,47 +333,6 @@ def verify():
    wait_for_condition(verify)


-def test_parent_task_id_non_concurrent_actor(shutdown_only):


For reviewer: See #32157 for more details about the context.

kevin85421 · 2025-03-17T06:08:17Z

[Done] Blocked by #51414

When we send a request to `/api/actors/kill` on the Ray dashboard with `force_kill=false`, the actor process will not actually be killed. See the following "Reproduction" section for more details. When `force_kill` is `false`, the function [CoreWorker::Exit](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1167) will post `drain_references_callback` to `task_execution_service_`. The callback function includes several parts, such as: * [Disconnect](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1078): * Based on my observation and experiments, no further `KillActor` RPCs are received after `local_raylet_client_->Disconnect` is called. * [Shutdown](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1010) * [task_receiver_->Stop()](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1028) will block until all running Ray tasks in the thread pools finish. * Non-threaded actor * In the following "Reproduction" section, the first request (`force_kill=false`) doesn't kill the process because `task_receiver_->Stop()` blocks execution, waiting for the actor task sleep to finish—which is an infinite loop. The second request (`force_kill=true`) kills the process because `drain_references_callback` was not called because both `drain_references_callback` and Ray task execution (i.e. `SleepActor.sleep` in this case) are submitted to `task_execution_service_` and both are executing on the main thread. That is, `Disconnect` is not called, so the process can still receive the force-kill `KillActor` RPC. * Threaded actor * The first request (`force_kill=false`) doesn't kill the process because `task_receiver_->Stop()` blocks execution, waiting for the actor task sleep to finish—which is an infinite loop. The second request (`force_kill=true`) doesn't kill the process because it cannot receive the force-kill `KillActor` RPC, as `Disconnect` was called after the first request. ### Solution This PR focuses on ensuring that core worker processes can still receive the `KillActor` RPC even when the previous RPC is blocked due to `task_receiver_->Stop()`. That is, the force-kill `KillActor` RPC can kill the process in the following "Reproduction" section. The solution is to ensure that `CoreWorker::Exit` does not block before calling `Disconnect`. Hence, this PR moves `task_receiver_->Stop()` to the very beginning of `drain_references_callback` instead of calling it after `Disconnect`. Possible followups: * Add a check to determine whether the concurrency group manager has already been stopped, and explicitly throw an exception if `Stop` is called twice. * Add timeout for `CoreWorker::Exit` to avoid blocking forever for long-running Ray tasks. ### Reproduction * `test.py` ```python import ray import time @ray.remote(concurrency_groups={"io": 1}) class SleepActor: def sleep(self): print("Sleep") while True: time.sleep(10) def echo(self, val): return val sa = SleepActor.remote() ray.get(sa.sleep.remote()) ``` * Reproduction ```sh ray start --head --include-dashboard=True --num-cpus=1 python3 test.py # Find the actor ID created by `test.py` ray list actors # Send a request to Ray dashboard `/api/actors/kill` with `force_kill=false`. curl -X GET -w "%{http_code}" "http://localhost:8265/api/actors/kill?actor_id=${YOUR_ACTOR_ID}&force_kill=false" # [Example output] {"result": true, "msg": "Requested actor with id ac24dd03dbfcb8ad02b5ceba01000000 to terminate. It will exit once running tasks complete", "data": {}}200% # [Example output of test.py] ray.exceptions.ActorDiedError: The actor died unexpectedly before finishing this task. class_name: SleepActor actor_id: 545606bd4458db0326a55b9701000000 pid: 2334580 namespace: 7b9c4d27-b70b-46d6-ad93-26f0da3e89f4 ip: 172.31.13.10 The actor is dead because its worker process has died. Worker exit type: INTENDED_SYSTEM_EXIT Worker exit detail: Worker exits because the actor is killed. The actor is dead because it was killed by `ray.kill`. # The actor is dead from the output of `ray list actors` ray list actors # The Ray Actor process is still running! ps aux | grep "ray::" # [Example output] ubuntu 2332284 0.5 0.1 19396584 73628 pts/12 SNl 05:02 0:01 ray::SleepActor.sleep # Force kill the actor curl -X GET -w "%{http_code}" "http://localhost:8265/api/actors/kill?actor_id=${YOUR_ACTOR_ID}&force_kill=false" # [Example output] {"result": true, "msg": "Force killed actor with id 545606bd4458db0326a55b9701000000", "data": {}}200% # The Ray Actor process is still running! ps aux | grep "ray::" ``` ## Related issue number #51058 Signed-off-by: kaihsun <kaihsun@anyscale.com>

…tmux-0-ray6

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 · 2025-03-18T20:26:10Z

python/ray/tests/test_actor.py

@@ -1046,7 +1046,6 @@ def get_value(self):
    )


-@pytest.mark.parametrize("enable_concurrency_group", [True, False])


For reviewer: This was added by #51203 to test that the tasks are not running on the main thread. Therefore, we can remove it for now.

kevin85421 · 2025-03-18T21:22:51Z

python/ray/tests/test_task_events_2.py

@@ -1118,90 +1118,13 @@ def verify_tasks(expected_tasks_cnt: Dict[str, int]):
    reason="setproctitle has different definitions of `title` on different OSes",
 )
 class TestIsActorTaskRunning:
-    def test_main_thread_short_comm(self, ray_start_regular):


For reviewer: See #51158 for more details. The Ray task is no longer running on the main thread, so these tests are not needed.

kevin85421 · 2025-03-18T21:30:21Z

python/ray/dashboard/modules/snapshot/tests/test_actors.py

@@ -50,16 +50,13 @@ def _kill_actor_using_dashboard_gcs(
    return resp_json


-@pytest.mark.parametrize("enable_concurrency_group", [False, True])


For reviewer: This is added by #51414. It's no longer needed.

Signed-off-by: kaihsun <kaihsun@anyscale.com>

…ain thread (#51516) Currently, some tests in `scheduling_queue_test.cc` rely on the behavior that both actor tasks and control plane logic execute on the same thread. This causes #51058 CI failures. This PR causes these tests to always run Ray actor tasks on the default executor instead of on the main thread. 1. If a test wants to check the status after all submitted tasks have finished, we call `default_executor->Join();` to wait for them to complete. 2. If the test wants to check the status after a part of submitted tasks have finished, we use `WaitForCondition` instead. --------- Signed-off-by: kaihsun <kaihsun@anyscale.com>

…tmux-0-ray6

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 · 2025-03-20T19:17:00Z

src/ray/core_worker/test/direct_actor_transport_test.cc

@@ -904,30 +905,32 @@ TEST_F(TaskReceiverTest, TestNewTaskFromDifferentWorker) {
  {
    auto request =
        CreatePushTaskRequestHelper(actor_id, 0, worker_id, caller_id, curr_timestamp);
-    rpc::PushTaskReply reply;


If we don't use a pointer, the stack will have been freed by the time the executor tries to access it.

If I use a smart pointer, I need to update the function signature.

chatted with @dentiny: use scope exit (abseil cleanup) or shared pointer

kevin85421 · 2025-03-21T06:55:02Z

[Done] Blocked by #51582

When we send a request to `/api/actors/kill` on the Ray dashboard with `force_kill=false`, the actor process will not actually be killed. See the following "Reproduction" section for more details. When `force_kill` is `false`, the function [CoreWorker::Exit](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1167) will post `drain_references_callback` to `task_execution_service_`. The callback function includes several parts, such as: * [Disconnect](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1078): * Based on my observation and experiments, no further `KillActor` RPCs are received after `local_raylet_client_->Disconnect` is called. * [Shutdown](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1010) * [task_receiver_->Stop()](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1028) will block until all running Ray tasks in the thread pools finish. * Non-threaded actor * In the following "Reproduction" section, the first request (`force_kill=false`) doesn't kill the process because `task_receiver_->Stop()` blocks execution, waiting for the actor task sleep to finish—which is an infinite loop. The second request (`force_kill=true`) kills the process because `drain_references_callback` was not called because both `drain_references_callback` and Ray task execution (i.e. `SleepActor.sleep` in this case) are submitted to `task_execution_service_` and both are executing on the main thread. That is, `Disconnect` is not called, so the process can still receive the force-kill `KillActor` RPC. * Threaded actor * The first request (`force_kill=false`) doesn't kill the process because `task_receiver_->Stop()` blocks execution, waiting for the actor task sleep to finish—which is an infinite loop. The second request (`force_kill=true`) doesn't kill the process because it cannot receive the force-kill `KillActor` RPC, as `Disconnect` was called after the first request. ### Solution This PR focuses on ensuring that core worker processes can still receive the `KillActor` RPC even when the previous RPC is blocked due to `task_receiver_->Stop()`. That is, the force-kill `KillActor` RPC can kill the process in the following "Reproduction" section. The solution is to ensure that `CoreWorker::Exit` does not block before calling `Disconnect`. Hence, this PR moves `task_receiver_->Stop()` to the very beginning of `drain_references_callback` instead of calling it after `Disconnect`. Possible followups: * Add a check to determine whether the concurrency group manager has already been stopped, and explicitly throw an exception if `Stop` is called twice. * Add timeout for `CoreWorker::Exit` to avoid blocking forever for long-running Ray tasks. ### Reproduction * `test.py` ```python import ray import time @ray.remote(concurrency_groups={"io": 1}) class SleepActor: def sleep(self): print("Sleep") while True: time.sleep(10) def echo(self, val): return val sa = SleepActor.remote() ray.get(sa.sleep.remote()) ``` * Reproduction ```sh ray start --head --include-dashboard=True --num-cpus=1 python3 test.py # Find the actor ID created by `test.py` ray list actors # Send a request to Ray dashboard `/api/actors/kill` with `force_kill=false`. curl -X GET -w "%{http_code}" "http://localhost:8265/api/actors/kill?actor_id=${YOUR_ACTOR_ID}&force_kill=false" # [Example output] {"result": true, "msg": "Requested actor with id ac24dd03dbfcb8ad02b5ceba01000000 to terminate. It will exit once running tasks complete", "data": {}}200% # [Example output of test.py] ray.exceptions.ActorDiedError: The actor died unexpectedly before finishing this task. class_name: SleepActor actor_id: 545606bd4458db0326a55b9701000000 pid: 2334580 namespace: 7b9c4d27-b70b-46d6-ad93-26f0da3e89f4 ip: 172.31.13.10 The actor is dead because its worker process has died. Worker exit type: INTENDED_SYSTEM_EXIT Worker exit detail: Worker exits because the actor is killed. The actor is dead because it was killed by `ray.kill`. # The actor is dead from the output of `ray list actors` ray list actors # The Ray Actor process is still running! ps aux | grep "ray::" # [Example output] ubuntu 2332284 0.5 0.1 19396584 73628 pts/12 SNl 05:02 0:01 ray::SleepActor.sleep # Force kill the actor curl -X GET -w "%{http_code}" "http://localhost:8265/api/actors/kill?actor_id=${YOUR_ACTOR_ID}&force_kill=false" # [Example output] {"result": true, "msg": "Force killed actor with id 545606bd4458db0326a55b9701000000", "data": {}}200% # The Ray Actor process is still running! ps aux | grep "ray::" ``` ## Related issue number ray-project#51058 Signed-off-by: kaihsun <kaihsun@anyscale.com>

…tmux-0-ray6

When we send a request to `/api/actors/kill` on the Ray dashboard with `force_kill=false`, the actor process will not actually be killed. See the following "Reproduction" section for more details. When `force_kill` is `false`, the function [CoreWorker::Exit](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1167) will post `drain_references_callback` to `task_execution_service_`. The callback function includes several parts, such as: * [Disconnect](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1078): * Based on my observation and experiments, no further `KillActor` RPCs are received after `local_raylet_client_->Disconnect` is called. * [Shutdown](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1010) * [task_receiver_->Stop()](https://github.com/ray-project/ray/blob/6ed8cfd08927ec30dc4d3bf777cda50b57bab28f/src/ray/core_worker/core_worker.cc#L1028) will block until all running Ray tasks in the thread pools finish. * Non-threaded actor * In the following "Reproduction" section, the first request (`force_kill=false`) doesn't kill the process because `task_receiver_->Stop()` blocks execution, waiting for the actor task sleep to finish—which is an infinite loop. The second request (`force_kill=true`) kills the process because `drain_references_callback` was not called because both `drain_references_callback` and Ray task execution (i.e. `SleepActor.sleep` in this case) are submitted to `task_execution_service_` and both are executing on the main thread. That is, `Disconnect` is not called, so the process can still receive the force-kill `KillActor` RPC. * Threaded actor * The first request (`force_kill=false`) doesn't kill the process because `task_receiver_->Stop()` blocks execution, waiting for the actor task sleep to finish—which is an infinite loop. The second request (`force_kill=true`) doesn't kill the process because it cannot receive the force-kill `KillActor` RPC, as `Disconnect` was called after the first request. ### Solution This PR focuses on ensuring that core worker processes can still receive the `KillActor` RPC even when the previous RPC is blocked due to `task_receiver_->Stop()`. That is, the force-kill `KillActor` RPC can kill the process in the following "Reproduction" section. The solution is to ensure that `CoreWorker::Exit` does not block before calling `Disconnect`. Hence, this PR moves `task_receiver_->Stop()` to the very beginning of `drain_references_callback` instead of calling it after `Disconnect`. Possible followups: * Add a check to determine whether the concurrency group manager has already been stopped, and explicitly throw an exception if `Stop` is called twice. * Add timeout for `CoreWorker::Exit` to avoid blocking forever for long-running Ray tasks. ### Reproduction * `test.py` ```python import ray import time @ray.remote(concurrency_groups={"io": 1}) class SleepActor: def sleep(self): print("Sleep") while True: time.sleep(10) def echo(self, val): return val sa = SleepActor.remote() ray.get(sa.sleep.remote()) ``` * Reproduction ```sh ray start --head --include-dashboard=True --num-cpus=1 python3 test.py # Find the actor ID created by `test.py` ray list actors # Send a request to Ray dashboard `/api/actors/kill` with `force_kill=false`. curl -X GET -w "%{http_code}" "http://localhost:8265/api/actors/kill?actor_id=${YOUR_ACTOR_ID}&force_kill=false" # [Example output] {"result": true, "msg": "Requested actor with id ac24dd03dbfcb8ad02b5ceba01000000 to terminate. It will exit once running tasks complete", "data": {}}200% # [Example output of test.py] ray.exceptions.ActorDiedError: The actor died unexpectedly before finishing this task. class_name: SleepActor actor_id: 545606bd4458db0326a55b9701000000 pid: 2334580 namespace: 7b9c4d27-b70b-46d6-ad93-26f0da3e89f4 ip: 172.31.13.10 The actor is dead because its worker process has died. Worker exit type: INTENDED_SYSTEM_EXIT Worker exit detail: Worker exits because the actor is killed. The actor is dead because it was killed by `ray.kill`. # The actor is dead from the output of `ray list actors` ray list actors # The Ray Actor process is still running! ps aux | grep "ray::" # [Example output] ubuntu 2332284 0.5 0.1 19396584 73628 pts/12 SNl 05:02 0:01 ray::SleepActor.sleep # Force kill the actor curl -X GET -w "%{http_code}" "http://localhost:8265/api/actors/kill?actor_id=${YOUR_ACTOR_ID}&force_kill=false" # [Example output] {"result": true, "msg": "Force killed actor with id 545606bd4458db0326a55b9701000000", "data": {}}200% # The Ray Actor process is still running! ps aux | grep "ray::" ``` ## Related issue number ray-project#51058 Signed-off-by: kaihsun <kaihsun@anyscale.com> Signed-off-by: Dhakshin Suriakannu <d_suriakannu@apple.com>

…ain thread (ray-project#51516) Currently, some tests in `scheduling_queue_test.cc` rely on the behavior that both actor tasks and control plane logic execute on the same thread. This causes ray-project#51058 CI failures. This PR causes these tests to always run Ray actor tasks on the default executor instead of on the main thread. 1. If a test wants to check the status after all submitted tasks have finished, we call `default_executor->Join();` to wait for them to complete. 2. If the test wants to check the status after a part of submitted tasks have finished, we use `WaitForCondition` instead. --------- Signed-off-by: kaihsun <kaihsun@anyscale.com> Signed-off-by: Dhakshin Suriakannu <d_suriakannu@apple.com>

kevin85421 commented Mar 4, 2025

View reviewed changes

kevin85421 assigned edoakes Mar 4, 2025

kevin85421 marked this pull request as ready for review March 4, 2025 17:07

kevin85421 marked this pull request as draft March 4, 2025 17:11

kevin85421 changed the title ~~[core] Always create a default executor~~ [WIP: not ready for review][core] Always create a default executor Mar 4, 2025

kevin85421 mentioned this pull request Mar 7, 2025

[core] Use the correct way to check whether an actor task is running or not #51158

Merged

8 tasks

kevin85421 added 2 commits March 8, 2025 04:57

create default executor

a5d9008

Signed-off-by: kaihsun <kaihsun@anyscale.com>

remove outdated tests

2ff306f

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 force-pushed the 20250303-devbox1-tmux-0-ray6 branch from 937e314 to 2ff306f Compare March 8, 2025 07:40

kevin85421 changed the title ~~[WIP: not ready for review][core] Always create a default executor~~ [core] Always create a default executor Mar 8, 2025

kevin85421 mentioned this pull request Mar 8, 2025

[core][dashboard] Lazy init NotifyQueue #51178

Merged

8 tasks

Merge remote-tracking branch 'upstream/master' into 20250303-devbox1-…

c6f34b4

…tmux-0-ray6

kevin85421 added the go add ONLY when ready to merge, run all tests label Mar 9, 2025

kevin85421 mentioned this pull request Mar 10, 2025

[core] Acquire GIL before calling PyGILState_Release #51203

Merged

8 tasks

kevin85421 added 2 commits March 10, 2025 15:44

Merge remote-tracking branch 'upstream/master' into 20250303-devbox1-…

ce34f93

…tmux-0-ray6

fix

5824dd2

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 commented Mar 13, 2025

View reviewed changes

This was referenced Mar 13, 2025

[core] Add the definition of parent_task_id in protobuf #51328

Merged

[core][1/N] Fix KillActor RPC for threaded actors #51414

Merged

kevin85421 added 3 commits March 18, 2025 19:55

Merge remote-tracking branch 'upstream/master' into 20250303-devbox1-…

f7ef930

…tmux-0-ray6

remove

db3a96c

Signed-off-by: kaihsun <kaihsun@anyscale.com>

remove

c4a3967

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 commented Mar 18, 2025

View reviewed changes

remove

5ffc9d6

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 mentioned this pull request Mar 19, 2025

[core] Avoid test logic that relies on actor tasks executing on the main thread #51516

Merged

8 tasks

kevin85421 added 2 commits March 19, 2025 08:26

update

7e5d446

Signed-off-by: kaihsun <kaihsun@anyscale.com>

Merge branch 'master' into 20250303-devbox1-tmux-0-ray6

4e18473

Merge remote-tracking branch 'upstream/master' into 20250303-devbox1-…

ec149c0

…tmux-0-ray6

kevin85421 force-pushed the 20250303-devbox1-tmux-0-ray6 branch from 4d8c918 to ec149c0 Compare March 20, 2025 16:49

update

6b0eb09

Signed-off-by: kaihsun <kaihsun@anyscale.com>

kevin85421 commented Mar 20, 2025

View reviewed changes

kevin85421 mentioned this pull request Mar 21, 2025

[core] Threaded actors get stuck forever if they receive two exit signals #51582

Merged

8 tasks

Merge remote-tracking branch 'upstream/master' into 20250303-devbox1-…

4333e84

…tmux-0-ray6

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[core] Always create a default executor #51058

[core] Always create a default executor #51058

kevin85421 commented Mar 4, 2025 •

edited

Loading

kevin85421 Mar 4, 2025

kevin85421 commented Mar 4, 2025

kevin85421 commented Mar 6, 2025 •

edited

Loading

kevin85421 commented Mar 9, 2025

kevin85421 commented Mar 9, 2025 •

edited

Loading

kevin85421 Mar 13, 2025 •

edited

Loading

kevin85421 commented Mar 17, 2025 •

edited

Loading

kevin85421 Mar 18, 2025 •

edited

Loading

kevin85421 Mar 18, 2025 •

edited

Loading

kevin85421 Mar 18, 2025 •

edited

Loading

kevin85421 Mar 20, 2025

kevin85421 Mar 25, 2025

kevin85421 commented Mar 21, 2025 •

edited

Loading

	if (is_asyncio_) {
	// Process async actor task.
	auto fiber = fiber_state_manager_->GetExecutor(request.ConcurrencyGroupName(),
	request.FunctionDescriptor());
	fiber->EnqueueFiber([this, request, task_id]() mutable {
	AcceptRequestOrRejectIfCanceled(task_id, request);
	});
	} else {
	// Process actor tasks.
	RAY_CHECK(pool_manager_ != nullptr);
	auto pool = pool_manager_->GetExecutor(request.ConcurrencyGroupName(),
	request.FunctionDescriptor());
	if (pool == nullptr) {
	AcceptRequestOrRejectIfCanceled(task_id, request);
	} else {
	pool->Post([this, request, task_id]() mutable {
	AcceptRequestOrRejectIfCanceled(task_id, request);
	});
	}
	}

		@@ -333,47 +333,6 @@ def verify():
		wait_for_condition(verify)


		def test_parent_task_id_non_concurrent_actor(shutdown_only):

		@@ -1046,7 +1046,6 @@ def get_value(self):
		)


		@pytest.mark.parametrize("enable_concurrency_group", [True, False])

		@@ -50,16 +50,13 @@ def _kill_actor_using_dashboard_gcs(
		return resp_json


		@pytest.mark.parametrize("enable_concurrency_group", [False, True])

[core] Always create a default executor #51058

Are you sure you want to change the base?

[core] Always create a default executor #51058

Conversation

kevin85421 commented Mar 4, 2025 • edited Loading

Why are these changes needed?

Related issue number

Checks

kevin85421 Mar 4, 2025

Choose a reason for hiding this comment

kevin85421 commented Mar 4, 2025

kevin85421 commented Mar 6, 2025 • edited Loading

kevin85421 commented Mar 9, 2025

kevin85421 commented Mar 9, 2025 • edited Loading

kevin85421 Mar 13, 2025 • edited Loading

Choose a reason for hiding this comment

kevin85421 commented Mar 17, 2025 • edited Loading

kevin85421 Mar 18, 2025 • edited Loading

Choose a reason for hiding this comment

kevin85421 Mar 18, 2025 • edited Loading

Choose a reason for hiding this comment

kevin85421 Mar 18, 2025 • edited Loading

Choose a reason for hiding this comment

kevin85421 Mar 20, 2025

Choose a reason for hiding this comment

kevin85421 Mar 25, 2025

Choose a reason for hiding this comment

kevin85421 commented Mar 21, 2025 • edited Loading

kevin85421 commented Mar 4, 2025 •

edited

Loading

kevin85421 commented Mar 6, 2025 •

edited

Loading

kevin85421 commented Mar 9, 2025 •

edited

Loading

kevin85421 Mar 13, 2025 •

edited

Loading

kevin85421 commented Mar 17, 2025 •

edited

Loading

kevin85421 Mar 18, 2025 •

edited

Loading

kevin85421 Mar 18, 2025 •

edited

Loading

kevin85421 Mar 18, 2025 •

edited

Loading

kevin85421 commented Mar 21, 2025 •

edited

Loading