def test_wait_failure_recovery_2(workflow_start_regular_shared):
# Test failing "workflow.wait" and its input steps.
@workflow.step
def sleep_identity(x: int):
# block the step by a global mark
while not utils.check_global_mark():
time.sleep(0.1)
time.sleep(x)
return x
@workflow.step
def identity(x):
return x
ws = [
sleep_identity.step(2),
sleep_identity.step(5),
sleep_identity.step(1),
]
w = workflow.wait(ws, num_returns=2, timeout=None)
utils.unset_global_mark()
_ = identity.step(w).run_async(workflow_id="wait_failure_recovery_2")
# wait util "workflow.wait" has been running
time.sleep(10)
workflow.cancel("wait_failure_recovery_2")
time.sleep(2)
utils.set_global_mark()
ready, unready = ray.get(workflow.resume("wait_failure_recovery_2"))
assert ready == [2, 1]
def test_wait_recovery_step_id(workflow_start_regular_shared):
# This test ensures workflow reuse the original directory and
# step id for "workflow.wait" during recovery.
@workflow.step
def identity(x: int):
# block the step by a global mark
assert utils.check_global_mark()
return x
w = workflow.wait([identity.step(42)], num_returns=1, timeout=None)
utils.unset_global_mark()
with pytest.raises(RaySystemError):
_ = w.run(workflow_id="test_wait_recovery_step_id")
utils.set_global_mark()
ready, unready = ray.get(workflow.resume("test_wait_recovery_step_id"))
assert ready == [42]
from ray.workflow import storage, workflow_storage
global_storage = storage.get_global_storage()
wf_storage = workflow_storage.WorkflowStorage("test_wait_recovery_step_id",
global_storage)
index = wf_storage.gen_step_id("workflow.wait")
# no new step id
assert index <= 1
def test_wait_for_multiple_events(workflow_start_regular_shared):
"""If a workflow has multiple event arguments, it should wait for them at the
same time.
"""
class EventListener1(workflow.EventListener):
async def poll_for_event(self):
utils.set_global_mark("listener1")
while not utils.check_global_mark("trigger_event"):
await asyncio.sleep(0.1)
return "event1"
class EventListener2(workflow.EventListener):
async def poll_for_event(self):
utils.set_global_mark("listener2")
while not utils.check_global_mark("trigger_event"):
await asyncio.sleep(0.1)
return "event2"
@ray.remote
def trivial_step(arg1, arg2):
return f"{arg1} {arg2}"
event1_promise = workflow.wait_for_event(EventListener1)
event2_promise = workflow.wait_for_event(EventListener2)
promise = workflow.create(trivial_step.bind(event1_promise,
event2_promise)).run_async()
while not (utils.check_global_mark("listener1")
and utils.check_global_mark("listener2")):
time.sleep(0.1)
utils.set_global_mark("trigger_event")
assert ray.get(promise) == "event1 event2"
def test_wait_failure_recovery_1(workflow_start_regular_shared):
# This tests that if a step using the output of "workflow.wait" as its
# input, it can be recovered after failure.
@workflow.step
def get_all(ready, unready):
return ready, unready
@workflow.step
def filter_all_2(wait_results):
assert wait_results[0] == [1, 3, 2]
# failure point
assert utils.check_global_mark()
ready, unready = wait_results
return get_all.step(ready, unready)
@workflow.step
def composite_2():
w = wait_multiple_steps.step()
return filter_all_2.step(w)
utils.unset_global_mark()
with pytest.raises(RaySystemError):
composite_2.step().run(workflow_id="wait_failure_recovery")
utils.set_global_mark()
ready, unready = ray.get(workflow.resume("wait_failure_recovery"))
assert ready == [1, 3, 2]
assert unready == [10, 12]
def test_recovery_simple(workflow_start_regular):
@ray.remote
def append1(x):
return x + "[append1]"
@ray.remote
def append2(x):
return x + "[append2]"
@ray.remote
def simple(x):
x = append1.bind(x)
y = the_failed_step.bind(x)
z = append2.bind(y)
return workflow.continuation(z)
utils.unset_global_mark()
workflow_id = "test_recovery_simple"
with pytest.raises(RaySystemError):
# internally we get WorkerCrashedError
workflow.create(simple.bind("x")).run(workflow_id=workflow_id)
assert workflow.get_status(
workflow_id) == workflow.WorkflowStatus.RESUMABLE
utils.set_global_mark()
output = workflow.resume(workflow_id)
assert ray.get(output) == "foo(x[append1])[append2]"
utils.unset_global_mark()
# resume from workflow output checkpoint
output = workflow.resume(workflow_id)
assert ray.get(output) == "foo(x[append1])[append2]"
def test_event_as_workflow(workflow_start_regular_shared):
class MyEventListener(workflow.EventListener):
async def poll_for_event(self):
while not utils.check_global_mark():
await asyncio.sleep(1)
utils.unset_global_mark()
promise = workflow.wait_for_event(MyEventListener).run_async("wf")
assert workflow.get_status("wf") == workflow.WorkflowStatus.RUNNING
utils.set_global_mark()
assert ray.get(promise) is None
def test_recovery_simple_1(workflow_start_regular):
utils.unset_global_mark()
workflow_id = "test_recovery_simple_1"
with pytest.raises(workflow.WorkflowExecutionError):
# internally we get WorkerCrashedError
workflow.run(the_failed_step.bind("x"), workflow_id=workflow_id)
assert workflow.get_status(workflow_id) == workflow.WorkflowStatus.FAILED
utils.set_global_mark()
assert workflow.resume(workflow_id) == "foo(x)"
utils.unset_global_mark()
# resume from workflow output checkpoint
assert workflow.resume(workflow_id) == "foo(x)"
def test_crash_during_event_checkpointing(workflow_start_regular_shared):
"""Ensure that if the cluster dies while the event is being checkpointed, we
properly re-poll for the event."""
from ray._private import storage
storage_uri = storage._storage_uri
"""Ensure that we don't re-call poll_for_event after `event_checkpointed`
returns, even after a crash."""
class MyEventListener(workflow.EventListener):
async def poll_for_event(self):
assert not utils.check_global_mark("committed")
if utils.check_global_mark("first"):
utils.set_global_mark("second")
utils.set_global_mark("first")
utils.set_global_mark("time_to_die")
while not utils.check_global_mark("resume"):
time.sleep(0.1)
async def event_checkpointed(self, event):
utils.set_global_mark("committed")
@ray.remote
def wait_then_finish(arg):
pass
event_promise = workflow.wait_for_event(MyEventListener)
workflow.run_async(wait_then_finish.bind(event_promise), workflow_id="workflow")
while not utils.check_global_mark("time_to_die"):
time.sleep(0.1)
assert utils.check_global_mark("first")
ray.shutdown()
subprocess.check_output(["ray", "stop", "--force"])
# Give the workflow some time to kill the cluster.
# time.sleep(3)
ray.init(num_cpus=4, storage=storage_uri)
workflow.init()
workflow.resume_async("workflow")
utils.set_global_mark("resume")
workflow.get_output("workflow")
assert utils.check_global_mark("second")
def test_recovery_complex(workflow_start_regular):
utils.unset_global_mark()
workflow_id = "test_recovery_complex"
with pytest.raises(RaySystemError):
# internally we get WorkerCrashedError
complex.step("x").run(workflow_id=workflow_id)
utils.set_global_mark()
output = workflow.resume(workflow_id)
r = "join(join(foo(x[append1]), [source1][append2]), join(x, [source1]))"
assert ray.get(output) == r
utils.unset_global_mark()
# resume from workflow output checkpoint
output = workflow.resume(workflow_id)
r = "join(join(foo(x[append1]), [source1][append2]), join(x, [source1]))"
assert ray.get(output) == r
def test_recovery_simple(workflow_start_regular):
utils.unset_global_mark()
workflow_id = "test_recovery_simple"
with pytest.raises(RaySystemError):
# internally we get WorkerCrashedError
simple.step("x").run(workflow_id=workflow_id)
assert workflow.get_status(
workflow_id) == workflow.WorkflowStatus.RESUMABLE
utils.set_global_mark()
output = workflow.resume(workflow_id)
assert ray.get(output) == "foo(x[append1])[append2]"
utils.unset_global_mark()
# resume from workflow output checkpoint
output = workflow.resume(workflow_id)
assert ray.get(output) == "foo(x[append1])[append2]"
def test_checkpoint_dag_recovery_partial(workflow_start_regular_shared):
utils.unset_global_mark()
start = time.time()
with pytest.raises(RaySystemError):
workflow.create(checkpoint_dag.bind(False)).run(
workflow_id="checkpoint_partial_recovery")
run_duration_partial = time.time() - start
utils.set_global_mark()
start = time.time()
recovered = ray.get(workflow.resume("checkpoint_partial_recovery"))
recover_duration_partial = time.time() - start
assert np.isclose(recovered, np.arange(SIZE).mean())
print(f"[partial] run_duration = {run_duration_partial}, "
f"recover_duration = {recover_duration_partial}")
def test_checkpoint_dag_recovery_whole(workflow_start_regular_shared):
utils.unset_global_mark()
start = time.time()
with pytest.raises(workflow.WorkflowExecutionError):
workflow.run(checkpoint_dag.bind(True),
workflow_id="checkpoint_whole_recovery")
run_duration_whole = time.time() - start
utils.set_global_mark()
start = time.time()
recovered = workflow.resume("checkpoint_whole_recovery")
recover_duration_whole = time.time() - start
assert np.isclose(recovered, np.arange(SIZE).mean())
print(f"[whole] run_duration = {run_duration_whole}, "
f"recover_duration = {recover_duration_whole}")
def test_recovery_complex(workflow_start_regular):
@ray.remote
def source1():
return "[source1]"
@ray.remote
def append1(x):
return x + "[append1]"
@ray.remote
def append2(x):
return x + "[append2]"
@ray.remote
def join(x, y):
return f"join({x}, {y})"
@ray.remote
def complex(x1):
x2 = source1.bind()
v = join.bind(x1, x2)
y = append1.bind(x1)
y = the_failed_step.bind(y)
z = append2.bind(x2)
u = join.bind(y, z)
return workflow.continuation(join.bind(u, v))
utils.unset_global_mark()
workflow_id = "test_recovery_complex"
with pytest.raises(workflow.WorkflowExecutionError):
# internally we get WorkerCrashedError
workflow.create(complex.bind("x")).run(workflow_id=workflow_id)
assert workflow.get_status(workflow_id) == workflow.WorkflowStatus.FAILED
utils.set_global_mark()
output = workflow.resume(workflow_id)
r = "join(join(foo(x[append1]), [source1][append2]), join(x, [source1]))"
assert ray.get(output) == r
utils.unset_global_mark()
# resume from workflow output checkpoint
output = workflow.resume(workflow_id)
r = "join(join(foo(x[append1]), [source1][append2]), join(x, [source1]))"
assert ray.get(output) == r
def test_checkpoint_dag_recovery_skip(workflow_start_regular_shared):
utils.unset_global_mark()
start = time.time()
with pytest.raises(workflow.WorkflowExecutionError):
workflow.create(
checkpoint_dag.options(**workflow.options(
checkpoint=False)).bind(False)).run(
workflow_id="checkpoint_skip_recovery")
run_duration_skipped = time.time() - start
utils.set_global_mark()
start = time.time()
recovered = ray.get(workflow.resume("checkpoint_skip_recovery"))
recover_duration_skipped = time.time() - start
assert np.isclose(recovered, np.arange(SIZE).mean())
print(f"[skipped] run_duration = {run_duration_skipped}, "
f"recover_duration = {recover_duration_skipped}")
def test_recovery_simple_2(workflow_start_regular):
@ray.remote
def simple(x):
return workflow.continuation(the_failed_step.bind(x))
utils.unset_global_mark()
workflow_id = "test_recovery_simple_2"
with pytest.raises(workflow.WorkflowExecutionError):
# internally we get WorkerCrashedError
workflow.create(simple.bind("x")).run(workflow_id=workflow_id)
assert workflow.get_status(workflow_id) == workflow.WorkflowStatus.FAILED
utils.set_global_mark()
output = workflow.resume(workflow_id)
assert ray.get(output) == "foo(x)"
utils.unset_global_mark()
# resume from workflow output checkpoint
output = workflow.resume(workflow_id)
assert ray.get(output) == "foo(x)"
async def event_checkpointed(self, event):
utils.set_global_mark("committed")
if utils.check_global_mark("first"):
utils.set_global_mark("second")
else:
utils.set_global_mark("first")
await asyncio.sleep(1000000)
async def poll_for_event(self):
assert not utils.check_global_mark("committed")
if utils.check_global_mark("first"):
utils.set_global_mark("second")
utils.set_global_mark("first")
utils.set_global_mark("time_to_die")
while not utils.check_global_mark("resume"):
time.sleep(0.1)
async def poll_for_event(self):
utils.set_global_mark("listener2")
while not utils.check_global_mark("trigger_event"):
await asyncio.sleep(0.1)
return "event2"
async def event_checkpointed(self, event):
utils.set_global_mark("committed")
def triggers_event():
utils.set_global_mark()
while not utils.check_global_mark("event_returning"):
time.sleep(0.1)
def never_ends(x):
utils.set_global_mark()
time.sleep(1000000)
return x
def triggers_event():
utils.set_global_mark()
def test_checkpoint_dag_recovery(workflow_start_regular):
utils.set_global_mark()
# warm up to ensure precise timing
for _ in range(3):
outputs = checkpoint_dag2.step(True).run()
assert np.isclose(outputs, 8388607.5)
utils.unset_global_mark()
start = time.time()
with pytest.raises(RaySystemError):
checkpoint_dag2.options(checkpoint=False).step(False).run(
workflow_id="checkpoint_skip2"
)
run_duration_skipped = time.time() - start
utils.set_global_mark()
start = time.time()
recovered = ray.get(workflow.resume("checkpoint_skip2"))
recover_duration_skipped = time.time() - start
assert np.isclose(recovered, 8388607.5)
utils.unset_global_mark()
start = time.time()
with pytest.raises(RaySystemError):
checkpoint_dag2.step(False).run(workflow_id="checkpoint_partial2")
run_duration_partial = time.time() - start
utils.set_global_mark()
start = time.time()
recovered = ray.get(workflow.resume("checkpoint_partial2"))
recover_duration_partial = time.time() - start
assert np.isclose(recovered, 8388607.5)
utils.unset_global_mark()
start = time.time()
with pytest.raises(RaySystemError):
checkpoint_dag2.step(True).run(workflow_id="checkpoint_whole2")
run_duration_whole = time.time() - start
utils.set_global_mark()
start = time.time()
recovered = ray.get(workflow.resume("checkpoint_whole2"))
recover_duration_whole = time.time() - start
assert np.isclose(recovered, 8388607.5)
print(
f"[skipped] run_duration = {run_duration_skipped}, "
f"recover_duration = {recover_duration_skipped}"
)
print(
f"[partial] run_duration = {run_duration_partial}, "
f"recover_duration = {recover_duration_partial}"
)
print(
f"[whole] run_duration = {run_duration_whole}, "
f"recover_duration = {recover_duration_whole}"
)
async def poll_for_event(self):
while not utils.check_global_mark():
await asyncio.sleep(0.1)
utils.set_global_mark("event_returning")
