def test_bounds_checking(self):
num_replicas = 10
max_replicas = 11
min_replicas = 9
config = AutoscalingConfig(
max_replicas=max_replicas,
min_replicas=min_replicas,
target_num_ongoing_requests_per_replica=100,
)
desired_num_replicas = calculate_desired_num_replicas(
autoscaling_config=config,
current_num_ongoing_requests=[150] * num_replicas)
assert desired_num_replicas == max_replicas
desired_num_replicas = calculate_desired_num_replicas(
autoscaling_config=config,
current_num_ongoing_requests=[50] * num_replicas)
assert desired_num_replicas == min_replicas
for i in range(50, 150):
desired_num_replicas = calculate_desired_num_replicas(
autoscaling_config=config,
current_num_ongoing_requests=[i] * num_replicas,
)
assert min_replicas <= desired_num_replicas <= max_replicas
def test_smoothing_factor(self):
config = AutoscalingConfig(min_replicas=0,
max_replicas=100,
target_num_ongoing_requests_per_replica=1,
smoothing_factor=0.5)
num_replicas = 10
num_ongoing_requests = [4.0] * num_replicas
desired_num_replicas = calculate_desired_num_replicas(
autoscaling_config=config,
current_num_ongoing_requests=num_ongoing_requests)
assert 24 <= desired_num_replicas <= 26 # 10 + 0.5 * (40 - 10) = 25
num_ongoing_requests = [0.25] * num_replicas
desired_num_replicas = calculate_desired_num_replicas(
autoscaling_config=config,
current_num_ongoing_requests=num_ongoing_requests)
assert 5 <= desired_num_replicas <= 8 # 10 + 0.5 * (2.5 - 10) = 6.25
def test_scale_down(self):
config = AutoscalingConfig(min_replicas=0,
max_replicas=100,
target_num_ongoing_requests_per_replica=1)
num_replicas = 10
num_ongoing_requests = [0.5] * num_replicas
desired_num_replicas = calculate_desired_num_replicas(
autoscaling_config=config,
current_num_ongoing_requests=num_ongoing_requests)
assert 4 <= desired_num_replicas <= 6 # 10 * 0.5 = 5
def test_scale_up(self):
config = AutoscalingConfig(min_replicas=0,
max_replicas=100,
target_num_ongoing_requests_per_replica=1)
num_replicas = 10
num_ongoing_requests = [2.0] * num_replicas
desired_num_replicas = calculate_desired_num_replicas(
autoscaling_config=config,
current_num_ongoing_requests=num_ongoing_requests)
assert 19 <= desired_num_replicas <= 21 # 10 * 2 = 20
def autoscale(self) -> None:
"""Update autoscaling deployments with calculated num_replicas."""
for deployment_name, (backend_info,
route_prefix) in self.list_deployments().items():
backend_config = backend_info.backend_config
autoscaling_config = backend_config.autoscaling_config
if autoscaling_config is None:
continue
replicas = self.backend_state_manager._backend_states[
deployment_name]._replicas
running_replicas = replicas.get([ReplicaState.RUNNING])
current_num_ongoing_requests = []
for replica in running_replicas:
replica_tag = replica.replica_tag
num_ongoing_requests = (
self.autoscaling_metrics_store.window_average(
replica_tag,
time.time() - autoscaling_config.look_back_period_s))
if num_ongoing_requests is not None:
current_num_ongoing_requests.append(num_ongoing_requests)
if len(current_num_ongoing_requests) == 0:
continue
new_backend_config = backend_config.copy()
new_backend_config.num_replicas = calculate_desired_num_replicas(
autoscaling_config, current_num_ongoing_requests)
replica_config = backend_info.replica_config
deployer_job_id = backend_info.deployer_job_id
backend_config_proto_bytes = new_backend_config.to_proto_bytes()
goal_id, updating = self.deploy(deployment_name,
backend_config_proto_bytes,
replica_config,
version=backend_info.version,
prev_version=backend_info.version,
route_prefix=route_prefix,
deployer_job_id=deployer_job_id)
