def __init__(self, vertex_config, model_name):
super().__init__()
handle_list = list()
for node_id in vertex_config.keys():
backend_config = vertex_config[node_id]
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(**backend_config)
if node_id == "prepoc":
min_img_size = 224
transform = transforms.Compose([
transforms.Resize(min_img_size),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
])
serve_benchmark.create_backend(Transform,
node_id,
transform,
backend_config=config)
elif node_id == "model":
serve_benchmark.create_backend(
PredictModelPytorch,
node_id,
model_name,
True,
backend_config=config,
)
serve_benchmark.link(node_id, node_id)
handle_list.append(serve_benchmark.get_handle(node_id))
self.chain_handle = ChainHandle(handle_list)
def test_e2e(serve_instance):
serve_benchmark.init() # so we have access to global state
serve_benchmark.create_endpoint("endpoint",
"/api",
methods=["GET", "POST"])
result = serve_benchmark.api._get_global_state().route_table.list_service()
assert result["/api"] == "endpoint"
retry_count = 5
timeout_sleep = 0.5
while True:
try:
resp = requests.get("http://127.0.0.1:8000/-/routes",
timeout=0.5).json()
assert resp == {"/api": ["endpoint", ["GET", "POST"]]}
break
except Exception as e:
time.sleep(timeout_sleep)
timeout_sleep *= 2
retry_count -= 1
if retry_count == 0:
assert False, ("Route table hasn't been updated after 3 tries."
"The latest error was {}").format(e)
def function(flask_request):
return {"method": flask_request.method}
serve_benchmark.create_backend(function, "echo:v1")
serve_benchmark.link("endpoint", "echo:v1")
resp = requests.get("http://127.0.0.1:8000/api").json()["method"]
assert resp == "GET"
resp = requests.post("http://127.0.0.1:8000/api").json()["method"]
assert resp == "POST"
def main(num_replicas, batch_size, method, filename):
for node_id in ["up", "down"]:
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(
max_batch_size=batch_size, num_replicas=num_replicas
)
serve_benchmark.create_backend(noop, node_id, backend_config=config)
serve_benchmark.link(node_id, node_id)
with serve_benchmark.using_router("up"):
up_handle = serve_benchmark.get_handle("up")
with serve_benchmark.using_router("down"):
down_handle = serve_benchmark.get_handle("down")
# start = time.perf_counter()
oids = []
if method == "chain":
# closed loop latencies
for i in range(num_queries):
r = up_handle.options(tracing_metadata={"pipeline-id": i}).remote(
data=image_data
)
oid = down_handle.options(
tracing_metadata={"pipeline-id": i}
).remote(
data=r # torch tensor
)
ray.wait([oid], 1)
elif method == "group":
res = [
up_handle.options(tracing_metadata={"pipeline-id": i}).remote(
sleep_time=0.01, data=image_data
)
for i in range(num_queries)
]
oids = [
down_handle.options(tracing_metadata={"pipeline-id": i}).remote(
sleep_time=0.02, data=d # torch tensor
)
for i, d in enumerate(res)
]
else:
raise RuntimeError("Unreachable")
# print(f"Submission time {time.perf_counter() - start}")
# end = time.perf_counter()
# duration = end - start
# qps = num_queries / duration
# print(f"Throughput {qps}")
trace_file = filename or tempfile.mkstemp(suffix=".json")[1]
with open(trace_file, "w") as f:
json.dump(serve_benchmark.get_trace(), f)
ray.timeline(f"ray_trace_{batch_size}-{method}.json")
print(f"Serve trace file written to {trace_file}")
def __init__(self, max_batch_size, pipeline_length):
self.plength = pipeline_length
self.handles = list()
for index in range(self.plength):
node_id = f"service-{index}"
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(
max_batch_size=max_batch_size, num_replicas=1)
serve_benchmark.create_backend(noop,
node_id,
backend_config=config)
serve_benchmark.link(node_id, node_id)
self.handles.append(serve_benchmark.get_handle(node_id))
def test_no_route(serve_instance):
serve_benchmark.create_endpoint("noroute-endpoint")
global_state = serve_benchmark.api._get_global_state()
result = global_state.route_table.list_service(include_headless=True)
assert result[NO_ROUTE_KEY] == ["noroute-endpoint"]
without_headless_result = global_state.route_table.list_service()
assert NO_ROUTE_KEY not in without_headless_result
def func(_, i=1):
return 1
serve_benchmark.create_backend(func, "backend:1")
serve_benchmark.link("noroute-endpoint", "backend:1")
service_handle = serve_benchmark.get_handle("noroute-endpoint")
result = ray.get(service_handle.remote(i=1))
assert result == 1
def test_scaling_replicas(serve_instance):
class Counter:
def __init__(self):
self.count = 0
def __call__(self, _):
self.count += 1
return self.count
serve_benchmark.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/-/routes").json()):
time.sleep(0.2)
b_config = BackendConfig(num_replicas=2)
serve_benchmark.create_backend(Counter,
"counter:v1",
backend_config=b_config)
serve_benchmark.link("counter", "counter:v1")
counter_result = []
for _ in range(10):
resp = requests.get("http://127.0.0.1:8000/increment").json()
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_benchmark.get_backend_config("counter:v1")
b_config.num_replicas = 1
serve_benchmark.set_backend_config("counter:v1", b_config)
counter_result = []
for _ in range(10):
resp = requests.get("http://127.0.0.1:8000/increment").json()
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_benchmark.accept_batch
def __call__(self, flask_request, temp=None):
batch_size = serve_benchmark.context.batch_size
return batch_size
serve_benchmark.create_endpoint("exception-test", "/noListReturned")
# set the max batch size
b_config = BackendConfig(max_batch_size=5)
serve_benchmark.create_backend(NoListReturned,
"exception:v1",
backend_config=b_config)
serve_benchmark.link("exception-test", "exception:v1")
handle = serve_benchmark.get_handle("exception-test")
with pytest.raises(ray.exceptions.RayTaskError):
assert ray.get(handle.remote(temp=1))
def test_batching(serve_instance):
class BatchingExample:
def __init__(self):
self.count = 0
@serve_benchmark.accept_batch
def __call__(self, flask_request, temp=None):
self.count += 1
batch_size = serve_benchmark.context.batch_size
return [self.count] * batch_size
serve_benchmark.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/-/routes").json()):
time.sleep(0.2)
# set the max batch size
b_config = BackendConfig(max_batch_size=5)
serve_benchmark.create_backend(BatchingExample,
"counter:v11",
backend_config=b_config)
serve_benchmark.link("counter1", "counter:v11")
future_list = []
handle = serve_benchmark.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 main(num_replicas, method):
for node_id in ["up", "down"]:
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(max_batch_size=1,
num_replicas=num_replicas)
serve_benchmark.create_backend(noop,
node_id,
backend_config=config)
serve_benchmark.link(node_id, node_id)
with serve_benchmark.using_router("up"):
up_handle = serve_benchmark.get_handle("up")
with serve_benchmark.using_router("down"):
down_handle = serve_benchmark.get_handle("down")
start = time.perf_counter()
oids = []
if method == "chain":
for i in range(num_queries):
r = up_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(sleep_time=0.01, data=image_data)
r = down_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(
sleep_time=0.02,
data=r # torch tensor
)
oids.append(r)
elif method == "group":
res = [
up_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(sleep_time=0.01, data=image_data)
for i in range(num_queries)
]
oids = [
down_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(
sleep_time=0.02,
data=d # torch tensor
) for i, d in enumerate(res)
]
else:
raise RuntimeError("Unreachable")
print(f"Submission time {time.perf_counter() - start}")
ray.wait(oids, len(oids))
end = time.perf_counter()
duration = end - start
qps = num_queries / duration
print(f"Throughput {qps}")
_, trace_file = tempfile.mkstemp(suffix=".json")
with open(trace_file, "w") as f:
json.dump(serve_benchmark.get_trace(), f)
print(f"trace file written to {trace_file}")