TaskCancellation

python/ray/__init__.py

@@ -66,6 +66,7 @@
     LOCAL_MODE,
     SCRIPT_MODE,
     WORKER_MODE,
+    cancel,
     connect,
     disconnect,
     get,
@@ -113,6 +114,7 @@
     "_config",
     "_get_runtime_context",
     "actor",
+    "cancel",
     "connect",
     "disconnect",
     "get",

python/ray/_raylet.pyx

@@ -17,6 +17,8 @@ import logging
 import os
 import pickle
 import sys
+import _thread
+import setproctitle
 
 from libc.stdint cimport (
     int32_t,
@@ -90,6 +92,7 @@ from ray.exceptions import (
     RayTaskError,
     ObjectStoreFullError,
     RayTimeoutError,
+    RayCancellationError
 )
 from ray.utils import decode
 import gc
@@ -452,14 +455,22 @@ cdef execute_task(
                 actor_title = "{}({}, {})".format(
                     class_name, repr(args), repr(kwargs))
                 core_worker.set_actor_title(actor_title.encode("utf-8"))
+            # Ensure no previous signals are still around
+            check_signals()
             # Execute the task.
             with ray.worker._changeproctitle(title, next_title):
                 with core_worker.profile_event(b"task:execute"):
                     task_exception = True
-                    outputs = function_executor(*args, **kwargs)
-                    task_exception = False
+                    try:
+                        outputs = function_executor(*args, **kwargs)
+                        task_exception = False
+                    except KeyboardInterrupt as e:
+                        raise RayCancellationError(
+                                core_worker.get_current_task_id())
                     if c_return_ids.size() == 1:
                         outputs = (outputs,)
+                # Ensure no signals are still around
+            check_signals()
             # Store the outputs in the object store.
             with core_worker.profile_event(b"task:store_outputs"):
                 core_worker.store_task_outputs(
@@ -551,6 +562,14 @@ cdef void async_plasma_callback(CObjectID object_id,
             event_handler._loop.call_soon_threadsafe(
                 event_handler._complete_future, obj_id)
 
+cdef c_bool kill_main_task() nogil:
+    with gil:
+        if setproctitle.getproctitle() != "ray::IDLE":
+            _thread.interrupt_main()
+            return True
+        return False
+
+
 cdef CRayStatus check_signals() nogil:
     with gil:
         try:
@@ -657,6 +676,7 @@ cdef class CoreWorker:
         options.ref_counting_enabled = True
         options.is_local_mode = local_mode
         options.num_workers = 1
+        options.kill_main = kill_main_task
 
         CCoreWorkerProcess.Initialize(options)
 
@@ -952,6 +972,17 @@ cdef class CoreWorker:
             check_status(CCoreWorkerProcess.GetCoreWorker().KillActor(
                   c_actor_id, True, no_reconstruction))
 
+    def kill_task(self, ObjectID object_id, c_bool force_kill):
+        cdef:
+            CObjectID c_object_id = object_id.native()
+            CRayStatus status = CRayStatus.OK()
+
+        status = CCoreWorkerProcess.GetCoreWorker().KillTask(
+                                            c_object_id, force_kill)
+
+        if not status.ok():
+            raise ValueError(status.message().decode())
+
     def resource_ids(self):
         cdef:
             ResourceMappingType resource_mapping = (

python/ray/exceptions.py

@@ -16,6 +16,22 @@ class RayConnectionError(RayError):
     pass
 
 
+class RayCancellationError(RayError):
+    """Raised when this task or a dependency is cancelled
+    Attributes:
+        task_id (TaskID): The TaskID of the function that was directly
+            cancelled.
+    """
+
+    def __init__(self, task_id=None):
+        self.task_id = task_id
+
+    def __str__(self):
+        if self.task_id is None:
+            return "This task or its dependency was cancelled by"
+        return "Task: " + str(self.task_id) + " was cancelled"
+
+
 class RayTaskError(RayError):
     """Indicates that a task threw an exception during execution.
 

python/ray/includes/libcoreworker.pxd

@@ -97,6 +97,7 @@ cdef extern from "ray/core_worker/core_worker.h" nogil:
         CRayStatus KillActor(
             const CActorID &actor_id, c_bool force_kill,
             c_bool no_reconstruction)
+        CRayStatus KillTask(const CObjectID &object_id, c_bool force_kill)
 
         unique_ptr[CProfileEvent] CreateProfileEvent(
             const c_string &event_type)
@@ -213,6 +214,7 @@ cdef extern from "ray/core_worker/core_worker.h" nogil:
         c_bool ref_counting_enabled
         c_bool is_local_mode
         int num_workers
+        (c_bool() nogil) kill_main
         CCoreWorkerOptions()
 
     cdef cppclass CCoreWorkerProcess "ray::CoreWorkerProcess":

python/ray/includes/unique_ids.pxd

@@ -122,6 +122,8 @@ cdef extern from "ray/common/id.h" namespace "ray" nogil:
         CTaskID ForNormalTask(CJobID job_id, CTaskID parent_task_id,
                               int64_t parent_task_counter)
 
+        CActorID ActorId() const
+
     cdef cppclass CObjectID" ray::ObjectID"(CBaseID[CObjectID]):
 
         @staticmethod

python/ray/includes/unique_ids.pxi

@@ -220,6 +220,9 @@ cdef class TaskID(BaseID):
     def is_nil(self):
         return self.data.IsNil()
 
+    def actor_id(self):
+        return ActorID(self.data.ActorId().Binary())
+
     cdef size_t hash(self):
         return self.data.Hash()
 

python/ray/serialization.py

@@ -12,6 +12,7 @@
     PlasmaObjectNotAvailable,
     RayTaskError,
     RayActorError,
+    RayCancellationError,
     RayWorkerError,
     UnreconstructableError,
 )
@@ -279,6 +280,8 @@ def _deserialize_object(self, data, metadata, object_id):
                 return RayWorkerError()
             elif error_type == ErrorType.Value("ACTOR_DIED"):
                 return RayActorError()
+            elif error_type == ErrorType.Value("TASK_CANCELLED"):
+                return RayCancellationError()
             elif error_type == ErrorType.Value("OBJECT_UNRECONSTRUCTABLE"):
                 return UnreconstructableError(ray.ObjectID(object_id.binary()))
             else:

python/ray/tests/BUILD

@@ -414,3 +414,11 @@ py_test(
     tags = ["exclusive"],
     deps = ["//:ray_lib"],
 )
+
+py_test(
+    name = "test_cancel",
+    size = "small",
+    srcs = ["test_cancel.py],
+    tags = ["exclusive"],
+    deps = ["//:ray_lib"],
+)

python/ray/tests/test_cancel.py

@@ -0,0 +1,191 @@
+import pytest
+import ray
+import random
+import sys
+import time
+from ray.exceptions import RayTaskError, RayTimeoutError, RayCancellationError
+from ray.test_utils import SignalActor
+
+
+@pytest.mark.parametrize("use_force", [True, False])
+def test_cancel_chain(ray_start_regular, use_force):
+    """A helper method for chain of events tests"""
+    signaler = SignalActor.remote()
+
+    @ray.remote
+    def wait_for(t):
+        return ray.get(t[0])
+
+    obj1 = wait_for.remote([signaler.wait.remote()])
+    obj2 = wait_for.remote([obj1])
+    obj3 = wait_for.remote([obj2])
+    obj4 = wait_for.remote([obj3])
+
+    assert len(ray.wait([obj1], timeout=.1)[0]) == 0
+    ray.cancel(obj1, use_force)
+    for ob in [obj1, obj2, obj3, obj4]:
+        with pytest.raises((RayTaskError, RayCancellationError)):
+            ray.get(ob)
+
+    signaler2 = SignalActor.remote()
+    obj1 = wait_for.remote([signaler2.wait.remote()])
+    obj2 = wait_for.remote([obj1])
+    obj3 = wait_for.remote([obj2])
+    obj4 = wait_for.remote([obj3])
+
+    assert len(ray.wait([obj3], timeout=.1)[0]) == 0
+    ray.cancel(obj3, use_force)
+    for ob in [obj3, obj4]:
+        with pytest.raises((RayTaskError, RayCancellationError)):
+            ray.get(ob)
+
+    with pytest.raises(RayTimeoutError):
+        ray.get(obj1, timeout=.1)
+
+    with pytest.raises(RayTimeoutError):
+        ray.get(obj2, timeout=.1)
+
+    signaler2.send.remote()
+    ray.get(obj1, timeout=.1)
+
+
+@pytest.mark.parametrize("use_force", [True, False])
+def test_cancel_multiple_dependents(ray_start_regular, use_force):
+    """A helper method for multiple waiters on events tests"""
+    signaler = SignalActor.remote()
+
+    @ray.remote
+    def wait_for(t):
+        return ray.get(t[0])
+
+    head = wait_for.remote([signaler.wait.remote()])
+    deps = []
+    for _ in range(3):
+        deps.append(wait_for.remote([head]))
+
+    assert len(ray.wait([head], timeout=.1)[0]) == 0
+    ray.cancel(head, use_force)
+    for d in deps:
+        with pytest.raises((RayTaskError, RayCancellationError)):
+            ray.get(d)
+
+    head2 = wait_for.remote([signaler.wait.remote()])
+
+    deps2 = []
+    for _ in range(3):
+        deps2.append(wait_for.remote([head]))
+
+    for d in deps2:
+        ray.cancel(d, use_force)
+
+    for d in deps2:
+        with pytest.raises((RayTaskError, RayCancellationError)):
+            ray.get(d)
+
+    signaler.send.remote()
+    ray.get(head2, timeout=1)
+
+
+@pytest.mark.parametrize("use_force", [True, False])
+def test_single_cpu_cancel(shutdown_only, use_force):
+    ray.init(num_cpus=1)
+    signaler = SignalActor.remote()
+
+    @ray.remote
+    def wait_for(t):
+        return ray.get(t[0])
+
+    obj1 = wait_for.remote([signaler.wait.remote()])
+    obj2 = wait_for.remote([obj1])
+    obj3 = wait_for.remote([obj2])
+    indep = wait_for.remote([signaler.wait.remote()])
+
+    assert len(ray.wait([obj3], timeout=.1)[0]) == 0
+    ray.cancel(obj3, use_force)
+    with pytest.raises((RayTaskError, RayCancellationError)):
+        ray.get(obj3, 0.1)
+
+    ray.cancel(obj1, use_force)
+
+    for d in [obj1, obj2]:
+        with pytest.raises((RayTaskError, RayCancellationError)):
+            ray.get(d)
+
+    signaler.send.remote()
+    ray.get(indep)
+
+
+@pytest.mark.parametrize("use_force", [True, False])
+def test_comprehensive(ray_start_regular, use_force):
+    signaler = SignalActor.remote()
+
+    @ray.remote
+    def wait_for(t):
+        ray.get(t[0])
+        return "Result"
+
+    @ray.remote
+    def combine(a, b):
+        return str(a) + str(b)
+
+    a = wait_for.remote([signaler.wait.remote()])
+    b = wait_for.remote([signaler.wait.remote()])
+    combo = combine.remote(a, b)
+    a2 = wait_for.remote([a])
+
+    assert len(ray.wait([a, b, a2, combo], timeout=1)[0]) == 0
+
+    ray.cancel(a, use_force)
+    with pytest.raises((RayTaskError, RayCancellationError)):
+        ray.get(a, 1)
+
+    with pytest.raises((RayTaskError, RayCancellationError)):
+        ray.get(a2, 1)
+
+    signaler.send.remote()
+
+    with pytest.raises((RayTaskError, RayCancellationError)):
+        ray.get(combo, 10)
+
+
+@pytest.mark.parametrize("use_force", [True, False])
+def test_stress(shutdown_only, use_force):
+    ray.init(num_cpus=1)
+
+    @ray.remote
+    def infinite_sleep(y):
+        if y:
+            while True:
+                time.sleep(1 / 10)
+
+    first = infinite_sleep.remote(True)
+
+    sleep_or_no = [random.randint(0, 1) for _ in range(100)]
+    tasks = [infinite_sleep.remote(i) for i in sleep_or_no]
+    cancelled = set()
+    for t in tasks:
+        if random.random() > 0.5:
+            ray.cancel(t, use_force)
+            cancelled.add(t)
+
+    ray.cancel(first, use_force)
+    cancelled.add(first)
+
+    for done in cancelled:
+        with pytest.raises((RayTaskError, RayCancellationError)):
+            ray.get(done, 10)
+
+    for indx in range(len(tasks)):
+        t = tasks[indx]
+        if sleep_or_no[indx]:
+            ray.cancel(t, use_force)
+            cancelled.add(t)
+        if t in cancelled:
+            with pytest.raises((RayTaskError, RayCancellationError)):
+                ray.get(t, 10)
+        else:
+            ray.get(t)
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main(["-v", __file__]))