def test_load_task(
printer: Callable,
client: cx.Client,
api: str,
load_config: Dict[str, Union[int, float]],
deploy_timeout: int = None,
retry_attempts: int = 0,
poll_sleep_seconds: int = 1,
api_config_name: str = "cortex.yaml",
):
jobs = load_config["jobs"]
concurrency = load_config["concurrency"]
submit_timeout = load_config["submit_timeout"]
workload_timeout = load_config["workload_timeout"]
api_dir = TEST_APIS_DIR / api
with open(str(api_dir / api_config_name)) as f:
api_specs = yaml.safe_load(f)
assert len(api_specs) == 1
api_name = api_specs[0]["name"]
client.create_api(api_spec=api_specs[0], project_dir=api_dir)
request_stopper = td.Event()
map_stopper = td.Event()
try:
assert endpoint_ready(
client=client, api_name=api_name,
timeout=deploy_timeout), f"api {api_name} not ready"
# give the operator time to start
time.sleep(1 * retry_attempts)
# submit jobs
printer(f"submitting {jobs} jobs concurrently")
job_specs = []
threads_futures = request_tasks_concurrently(client, api_name,
request_stopper,
concurrency, jobs,
job_specs)
assert wait_on_event(
request_stopper, submit_timeout
), f"{jobs} jobs couldn't be submitted in {submit_timeout}s"
check_futures_healthy(threads_futures)
wait_on_futures(threads_futures)
printer("waiting on the jobs")
job_ids = [job_spec.json()["job_id"] for job_spec in job_specs]
retrieve_results_concurrently(
client,
api_name,
concurrency,
map_stopper,
job_ids,
poll_sleep_seconds=poll_sleep_seconds,
timeout=workload_timeout,
)
except:
# best effort
try:
api_info = client.get_api(api_name)
# only get the last 10 job statuses
if "task_job_statuses" in api_info and len(
api_info["task_job_statuses"]) > 10:
api_info["task_job_statuses"] = api_info["task_job_statuses"][
-10:]
printer(json.dumps(api_info, indent=2))
except:
pass
raise
finally:
map_stopper.set()
delete_apis(client, [api_name])
def test_load_async(
printer: Callable,
client: cx.Client,
api: str,
load_config: Dict[str, Union[int, float]],
deploy_timeout: int = None,
poll_sleep_seconds: int = 1,
api_config_name: str = "cortex.yaml",
):
total_requests = load_config["total_requests"]
desired_replicas = load_config["desired_replicas"]
concurrency = load_config["concurrency"]
submit_timeout = load_config["submit_timeout"]
workload_timeout = load_config["workload_timeout"]
api_dir = TEST_APIS_DIR / api
with open(str(api_dir / api_config_name)) as f:
api_specs = yaml.safe_load(f)
assert len(api_specs) == 1
api_specs[0]["autoscaling"] = {
"min_replicas": desired_replicas,
"max_replicas": desired_replicas,
}
api_name = api_specs[0]["name"]
client.create_api(api_spec=api_specs[0], project_dir=api_dir)
request_stopper = td.Event()
map_stopper = td.Event()
responses: List[Dict[str, Any]] = []
try:
printer(f"getting {desired_replicas} replicas ready")
assert apis_ready(client=client,
api_names=[api_name],
timeout=deploy_timeout), f"api {api_name} not ready"
with open(str(api_dir / "sample.json")) as f:
payload = json.load(f)
printer("start making prediction requests concurrently")
threads_futures = make_requests_concurrently(
client,
api_name,
concurrency,
request_stopper,
responses=responses,
max_total_requests=total_requests,
payload=payload,
)
assert wait_on_event(
request_stopper, submit_timeout
), f"{total_requests} couldn't be submitted in {submit_timeout}s"
check_futures_healthy(threads_futures)
wait_on_futures(threads_futures)
printer("finished making prediction requests")
assert (
len(responses) == total_requests
), f"the submitted number of requests doesn't match the returned number of responses"
job_ids = []
for response in responses:
response_json = response.json()
assert "id" in response_json
job_ids.append(response_json["id"])
# assert the results
printer("start retrieving the async results concurrently")
results = []
retrieve_results_concurrently(
client,
api_name,
concurrency,
map_stopper,
job_ids,
results,
poll_sleep_seconds,
workload_timeout,
)
for request_id, json_result in results:
# validate keys are in the result json response
assert (
"id" in json_result
), f"id key was not present in result response (response: {json_result})"
assert (
"result" in json_result
), f"result key was not present in result response (response: {json_result})"
assert (
"timestamp" in json_result
), f"timestamp key was not present in result response (response: {json_result})"
# validate result json response has valid values
assert (
json_result["id"] == request_id
), f"result 'id' and request 'id' mismatch ({json_result['id']} != {request_id})"
assert (
json_result["status"] == "completed"
), f"async workload did not complete (response: {json_result})"
assert json_result[
"timestamp"] != "", "result 'timestamp' value was empty"
assert json_result[
"result"] != "", "result 'result' value was empty"
except:
# best effort
try:
api_info = client.get_api(api_name)
printer(json.dumps(api_info, indent=2))
except:
pass
raise
finally:
if "results" in vars() and len(results) < total_requests:
printer(
f"{len(results)}/{total_requests} have been successfully retrieved"
)
map_stopper.set()
delete_apis(client, [api_name])
def test_load_realtime(
printer: Callable,
client: cx.Client,
api: str,
load_config: Dict[str, Union[int, float]],
deploy_timeout: int = None,
api_config_name: str = "cortex.yaml",
):
total_requests = load_config["total_requests"]
desired_replicas = load_config["desired_replicas"]
concurrency = load_config["concurrency"]
min_rtt = load_config["min_rtt"]
max_rtt = load_config["max_rtt"]
avg_rtt = load_config["avg_rtt"]
avg_rtt_tolerance = load_config["avg_rtt_tolerance"]
status_code_timeout = load_config["status_code_timeout"]
api_dir = TEST_APIS_DIR / api
with open(str(api_dir / api_config_name)) as f:
api_specs = yaml.safe_load(f)
assert len(api_specs) == 1
api_specs[0]["autoscaling"] = {
"min_replicas": desired_replicas,
"max_replicas": desired_replicas,
}
api_name = api_specs[0]["name"]
client.create_api(api_spec=api_specs[0], project_dir=api_dir)
# controls the flow of requests
request_stopper = td.Event()
latencies: List[float] = []
try:
printer(f"getting {desired_replicas} replicas ready")
assert apis_ready(client=client,
api_names=[api_name],
timeout=deploy_timeout), f"api {api_name} not ready"
# give the APIs some time to prevent getting high latency spikes in the beginning
time.sleep(5)
with open(str(api_dir / "sample.json")) as f:
payload = json.load(f)
printer("start making requests concurrently")
threads_futures = make_requests_concurrently(
client,
api_name,
concurrency,
request_stopper,
latencies=latencies,
max_total_requests=total_requests,
payload=payload,
)
while not request_stopper.is_set():
current_min_rtt = min(latencies) if len(latencies) > 0 else min_rtt
assert (
current_min_rtt >= min_rtt
), f"min latency threshold hit; got {current_min_rtt}s, but the lowest accepted latency is {min_rtt}s"
current_max_rtt = max(latencies) if len(latencies) > 0 else max_rtt
assert (
current_max_rtt <= max_rtt
), f"max latency threshold hit; got {current_max_rtt}s, but the highest accepted latency is {max_rtt}s"
current_avg_rtt = sum(latencies) / len(latencies) if len(
latencies) > 0 else avg_rtt
assert (
current_avg_rtt > avg_rtt - avg_rtt_tolerance
and current_avg_rtt < avg_rtt + avg_rtt_tolerance
), f"avg latency ({current_avg_rtt}s) falls outside the expected range ({avg_rtt - avg_rtt_tolerance}s - {avg_rtt + avg_rtt_tolerance})"
api_info = client.get_api(api_name)
network_stats = api_info["metrics"]["network_stats"]
assert (
network_stats["code_4xx"] == 0
), f"detected 4xx response codes ({network_stats['code_4xx']}) in cortex get"
assert (
network_stats["code_5xx"] == 0
), f"detected 5xx response codes ({network_stats['code_5xx']}) in cortex get"
printer(
f"min RTT: {current_min_rtt} | max RTT: {current_max_rtt} | avg RTT: {current_avg_rtt} | requests: {network_stats['code_2xx']} (out of {total_requests})"
)
# check if the requesting threads are still healthy
# if not, they'll raise an exception
check_futures_healthy(threads_futures)
# don't stress the CPU too hard
time.sleep(1)
printer("verifying number of processed requests using the client")
assert api_requests(
client, api_name, total_requests, timeout=status_code_timeout
), f"the number of 2xx response codes for api {api_name} doesn't match the expected number {total_requests}"
except:
# best effort
try:
api_info = client.get_api(api_name)
printer(json.dumps(api_info, indent=2))
except:
pass
raise
finally:
request_stopper.set()
delete_apis(client, [api_name])
def test_autoscaling(
printer: Callable,
client: cx.Client,
apis: Dict[str, Any],
autoscaling_config: Dict[str, Union[int, float]],
deploy_timeout: int = None,
api_config_name: str = "cortex.yaml",
):
max_replicas = autoscaling_config["max_replicas"]
query_params = apis["query_params"]
# increase the concurrency by 1 to ensure we get max_replicas replicas
concurrency = max_replicas + 1
all_apis = [apis["primary"]] + apis["dummy"]
all_api_names = []
for api in all_apis:
api_dir = TEST_APIS_DIR / api
with open(str(api_dir / api_config_name)) as f:
api_specs = yaml.safe_load(f)
assert len(api_specs) == 1
api_specs[0]["autoscaling"] = {
"max_replicas": max_replicas,
"downscale_stabilization_period": "1m",
}
all_api_names.append(api_specs[0]["name"])
client.create_api(api_spec=api_specs[0], project_dir=api_dir)
primary_api_name = all_api_names[0]
autoscaling = client.get_api(primary_api_name)["spec"]["autoscaling"]
# controls the flow of requests
request_stopper = td.Event()
# determine upscale/downscale replica requests
current_replicas = 1 # starting number of replicas
test_timeout = 0 # measured in seconds
while current_replicas < max_replicas:
upscale_ceil = math.ceil(current_replicas *
autoscaling["max_upscale_factor"])
if upscale_ceil > current_replicas + 1:
current_replicas = upscale_ceil
else:
current_replicas += 1
if current_replicas > max_replicas:
current_replicas = max_replicas
test_timeout += int(autoscaling["upscale_stabilization_period"] /
(1000**3))
while current_replicas > 1:
downscale_ceil = math.ceil(current_replicas *
autoscaling["max_downscale_factor"])
if downscale_ceil < current_replicas - 1:
current_replicas = downscale_ceil
else:
current_replicas -= 1
test_timeout += int(autoscaling["downscale_stabilization_period"] /
(1000**3))
# add overhead to the test timeout to account for the process of downloading images or adding nodes to the cluster
test_timeout *= 2
try:
assert apis_ready(
client=client, api_names=all_api_names,
timeout=deploy_timeout), f"apis {all_api_names} not ready"
threads_futures = make_requests_concurrently(client,
primary_api_name,
concurrency,
request_stopper,
query_params=query_params)
test_start_time = time.time()
# upscale/downscale the api
printer(f"scaling up to {max_replicas} replicas")
while True:
assert api_updated(
client, primary_api_name, timeout=deploy_timeout
), "api didn't scale up to the desired number of replicas in time"
current_replicas = client.get_api(
primary_api_name)["status"]["replica_counts"]["requested"]
# stop the requests from being made
if current_replicas == max_replicas and not request_stopper.is_set(
):
printer(f"scaling back down to 1 replica")
request_stopper.set()
# check if the requesting threads are still healthy
# if not, they'll raise an exception
check_futures_healthy(threads_futures)
# check if the test is taking too much time
assert (
time.time() - test_start_time < test_timeout
), f"autoscaling test for api {primary_api_name} did not finish in {test_timeout}s; current number of replicas is {current_replicas}/{concurrency}"
# stop the test if it has finished
if current_replicas == 1 and request_stopper.is_set():
break
# add some delay to reduce the number of gets
time.sleep(1)
except:
# best effort
try:
api_info = client.get_api(primary_api_name)
printer(json.dumps(api_info, indent=2))
except:
pass
raise
finally:
request_stopper.set()
delete_apis(client, all_api_names)
def test_load_realtime(
printer: Callable,
client: cx.Client,
api: str,
load_config: Dict[str, Union[int, float]],
deploy_timeout: int = None,
api_config_name: str = "cortex_cpu.yaml",
node_groups: List[str] = [],
):
total_requests = load_config["total_requests"]
desired_replicas = load_config["desired_replicas"]
concurrency = load_config["concurrency"]
status_code_timeout = load_config["status_code_timeout"]
api_dir = TEST_APIS_DIR / api
with open(str(api_dir / api_config_name)) as f:
api_specs = yaml.safe_load(f)
assert len(api_specs) == 1
api_specs[0]["autoscaling"] = {
"min_replicas": desired_replicas,
"max_replicas": desired_replicas,
}
if len(node_groups) > 0:
api_specs[0]["node_groups"] = node_groups
api_name = api_specs[0]["name"]
client.deploy(api_spec=api_specs[0])
# controls the flow of requests
request_stopper = td.Event()
failed = False
try:
printer(f"getting {desired_replicas} replicas ready")
assert apis_ready(client=client,
api_names=[api_name],
timeout=deploy_timeout), f"api {api_name} not ready"
# give the APIs some time to prevent getting high latency spikes in the beginning
time.sleep(5)
with open(str(api_dir / "sample.json")) as f:
payload = json.load(f)
printer("start making requests concurrently")
threads_futures = make_requests_concurrently(
client,
api_name,
concurrency,
request_stopper,
max_total_requests=total_requests,
payload=payload,
)
while not request_stopper.is_set():
# check if the requesting threads are still healthy
# if not, they'll raise an exception
check_futures_healthy(threads_futures)
# don't stress the CPU too hard
time.sleep(1)
except:
# best effort
failed = True
try:
api_info = client.get_api(api_name)
printer(json.dumps(api_info, indent=2))
finally:
raise
finally:
request_stopper.set()
delete_apis(client, [api_name])
if failed:
time.sleep(30)
def test_load_realtime(
printer: Callable,
client: cx.Client,
api: str,
load_config: Dict[str, Union[int, float]],
deploy_timeout: int = None,
api_config_name: str = "cortex_cpu.yaml",
node_groups: List[str] = [],
):
total_requests = load_config["total_requests"]
desired_replicas = load_config["desired_replicas"]
concurrency = load_config["concurrency"]
status_code_timeout = load_config["status_code_timeout"]
api_dir = TEST_APIS_DIR / api
with open(str(api_dir / api_config_name)) as f:
api_specs = yaml.safe_load(f)
assert len(api_specs) == 1
api_specs[0]["autoscaling"] = {
"min_replicas": desired_replicas,
"max_replicas": desired_replicas,
}
if len(node_groups) > 0:
api_specs[0]["node_groups"] = node_groups
api_name = api_specs[0]["name"]
client.deploy(api_spec=api_specs[0])
# controls the flow of requests
request_stopper = td.Event()
failed = False
try:
printer(f"getting {desired_replicas} replicas ready")
assert apis_ready(client=client,
api_names=[api_name],
timeout=deploy_timeout), f"api {api_name} not ready"
network_stats = client.get_api(api_name)["metrics"]["network_stats"]
offset_2xx = network_stats["code_2xx"]
offset_4xx = network_stats["code_4xx"]
offset_5xx = network_stats["code_5xx"]
if offset_2xx is None:
offset_2xx = 0
if offset_4xx is None:
offset_4xx = 0
if offset_5xx is None:
offset_5xx = 0
# give the APIs some time to prevent getting high latency spikes in the beginning
time.sleep(5)
with open(str(api_dir / "sample.json")) as f:
payload = json.load(f)
printer("start making requests concurrently")
threads_futures = make_requests_concurrently(
client,
api_name,
concurrency,
request_stopper,
max_total_requests=total_requests,
payload=payload,
)
while not request_stopper.is_set():
api_info = client.get_api(api_name)
network_stats = api_info["metrics"]["network_stats"]
assert (
network_stats["code_4xx"] - offset_4xx == 0
), f"detected 4xx response codes ({network_stats['code_4xx'] - offset_4xx}) in cortex get"
assert (
network_stats["code_5xx"] - offset_5xx == 0
), f"detected 5xx response codes ({network_stats['code_5xx'] - offset_5xx}) in cortex get"
# check if the requesting threads are still healthy
# if not, they'll raise an exception
check_futures_healthy(threads_futures)
# don't stress the CPU too hard
time.sleep(1)
printer(
f"verifying number of processed requests ({total_requests}, with an offset of {offset_2xx}) using the client"
)
assert api_requests(
client,
api_name,
total_requests + offset_2xx,
timeout=status_code_timeout
), f"the number of 2xx response codes for api {api_name} doesn't match the expected number {total_requests}"
except:
# best effort
failed = True
try:
api_info = client.get_api(api_name)
printer(json.dumps(api_info, indent=2))
finally:
raise
finally:
request_stopper.set()
delete_apis(client, [api_name])
if failed:
time.sleep(30)