def test_batching(serve_instance):
class BatchingExample:
def __init__(self):
self.count = 0
@serve.accept_batch
def __call__(self, flask_request, temp=None):
self.count += 1
batch_size = serve.context.batch_size
return [self.count] * batch_size
serve.create_endpoint("counter1", "/increment")
# Keep checking the routing table until /increment is populated
while "/increment" not in requests.get("http://127.0.0.1:8000/").json():
time.sleep(0.2)
# set the max batch size
b_config = BackendConfig(max_batch_size=5)
serve.create_backend(BatchingExample,
"counter:v11",
backend_config=b_config)
serve.link("counter1", "counter:v11")
future_list = []
handle = serve.get_handle("counter1")
for _ in range(20):
f = handle.remote(temp=1)
future_list.append(f)
counter_result = ray.get(future_list)
# since count is only updated per batch of queries
# If there atleast one __call__ fn call with batch size greater than 1
# counter result will always be less than 20
assert max(counter_result) < 20
def test_not_killing_replicas(serve_instance):
class BatchSimple:
def __init__(self):
self.count = 0
@serve.accept_batch
def __call__(self, flask_request, temp=None):
batch_size = serve.context.batch_size
return [1] * batch_size
serve.create_endpoint("bsimple", "/bsimple")
b_config = BackendConfig(num_replicas=3, max_batch_size=2)
serve.create_backend(BatchSimple, "bsimple:v1", backend_config=b_config)
global_state = serve.api._get_global_state()
old_replica_tag_list = global_state.backend_table.list_replicas(
"bsimple:v1")
bnew_config = serve.get_backend_config("bsimple:v1")
# change the config
bnew_config.max_batch_size = 5
# set the config
serve.set_backend_config("bsimple:v1", bnew_config)
new_replica_tag_list = global_state.backend_table.list_replicas(
"bsimple:v1")
global_state.refresh_actor_handle_cache()
new_all_tag_list = list(global_state.actor_handle_cache.keys())
# the old and new replica tag list should be identical
# and should be subset of all_tag_list
assert set(old_replica_tag_list) <= set(new_all_tag_list)
assert set(old_replica_tag_list) == set(new_replica_tag_list)
def test_killing_replicas(serve_instance):
class Simple:
def __init__(self):
self.count = 0
def __call__(self, flask_request, temp=None):
return temp
serve.create_endpoint("simple", "/simple")
b_config = BackendConfig(num_replicas=3, num_cpus=2)
serve.create_backend(Simple, "simple:v1", backend_config=b_config)
global_state = serve.api._get_global_state()
old_replica_tag_list = global_state.backend_table.list_replicas(
"simple:v1")
bnew_config = serve.get_backend_config("simple:v1")
# change the config
bnew_config.num_cpus = 1
# set the config
serve.set_backend_config("simple:v1", bnew_config)
new_replica_tag_list = global_state.backend_table.list_replicas(
"simple:v1")
global_state.refresh_actor_handle_cache()
new_all_tag_list = list(global_state.actor_handle_cache.keys())
# the new_replica_tag_list must be subset of all_tag_list
assert set(new_replica_tag_list) <= set(new_all_tag_list)
# the old_replica_tag_list must not be subset of all_tag_list
assert not set(old_replica_tag_list) <= set(new_all_tag_list)
def test_scaling_replicas(serve_instance):
class Counter:
def __init__(self):
self.count = 0
def __call__(self, _):
self.count += 1
return self.count
serve.create_endpoint("counter", "/increment")
# Keep checking the routing table until /increment is populated
while "/increment" not in requests.get("http://127.0.0.1:8000/").json():
time.sleep(0.2)
b_config = BackendConfig(num_replicas=2)
serve.create_backend(Counter, "counter:v1", backend_config=b_config)
serve.link("counter", "counter:v1")
counter_result = []
for _ in range(10):
resp = requests.get("http://127.0.0.1:8000/increment").json()["result"]
counter_result.append(resp)
# If the load is shared among two replicas. The max result cannot be 10.
assert max(counter_result) < 10
b_config = serve.get_backend_config("counter:v1")
b_config.num_replicas = 1
serve.set_backend_config("counter:v1", b_config)
counter_result = []
for _ in range(10):
resp = requests.get("http://127.0.0.1:8000/increment").json()["result"]
counter_result.append(resp)
# Give some time for a replica to spin down. But majority of the request
# should be served by the only remaining replica.
assert max(counter_result) - min(counter_result) > 6
def test_batching_exception(serve_instance):
class NoListReturned:
def __init__(self):
self.count = 0
@serve.accept_batch
def __call__(self, flask_request, temp=None):
batch_size = serve.context.batch_size
return batch_size
serve.create_endpoint("exception-test", "/noListReturned")
# set the max batch size
b_config = BackendConfig(max_batch_size=5)
serve.create_backend(NoListReturned,
"exception:v1",
backend_config=b_config)
serve.link("exception-test", "exception:v1")
handle = serve.get_handle("exception-test")
with pytest.raises(ray.exceptions.RayTaskError):
assert ray.get(handle.remote(temp=1))
'info': {
"patient_name": "Adam",
"value": 0.0,
"vtype": "ECG"
}
}
# create ECG service
serve.create_endpoint("ECG")
# create data point service for hospital
serve.create_endpoint("hospital",
route="/hospital",
kwargs_creator=kwargs_creator)
# create backend for ECG
b_config = BackendConfig(num_replicas=1)
serve.create_backend(PytorchPredictorECG,
"PredictECG",
model,
cuda,
backend_config=b_config)
# link service and backend
serve.link("ECG", "PredictECG")
handle = serve.get_handle("ECG")
# prepare args for StorePatientData backend.
service_handles_dict = {"ECG": handle}
# do prediction after every 3750 queries.
num_queries_dict = {"ECG": 3750}
# Always keep num_replicas as 1 as this is a stateful Backend
# This backend will store all the patient's data and transfer
result = []
for flask_request in flask_request_list:
base_number = int(flask_request.args.get("base_number", "0"))
result.append(base_number)
return list(map(lambda x: x + self.increment, result))
else:
result = []
for b in base_number:
ans = b + self.increment
result.append(ans)
return result
serve.init(blocking=True)
serve.create_endpoint("magic_counter", "/counter", blocking=True)
b_config = BackendConfig(max_batch_size=5)
serve.create_backend(MagicCounter, "counter:v1", 42,
backend_config=b_config) # increment=42
serve.link("magic_counter", "counter:v1")
print("Sending ten queries via HTTP")
for i in range(10):
url = "http://127.0.0.1:8000/counter?base_number={}".format(i)
print("> Pinging {}".format(url))
resp = requests.get(url).json()
print(pformat_color_json(resp))
time.sleep(0.2)
print("Sending ten queries via Python")
handle = serve.get_handle("magic_counter")
