def autoscale_marathon_instance(marathon_service_config, marathon_tasks,
mesos_tasks):
current_instances = marathon_service_config.get_instances()
if len(marathon_tasks) != current_instances:
write_to_log(
config=marathon_service_config,
line=
'Delaying scaling as marathon is either waiting for resources or is delayed'
)
return
autoscaling_params = marathon_service_config.get_autoscaling_params()
autoscaling_metrics_provider = get_service_metrics_provider(
autoscaling_params.pop(SERVICE_METRICS_PROVIDER_KEY))
autoscaling_decision_policy = get_decision_policy(
autoscaling_params.pop(DECISION_POLICY_KEY))
utilization = autoscaling_metrics_provider(
marathon_service_config=marathon_service_config,
marathon_tasks=marathon_tasks,
mesos_tasks=mesos_tasks,
**autoscaling_params)
error = get_error_from_utilization(
utilization=utilization,
setpoint=autoscaling_params.pop('setpoint'),
current_instances=current_instances,
)
zookeeper_path = compose_autoscaling_zookeeper_root(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
)
autoscaling_amount = autoscaling_decision_policy(
error=error,
min_instances=marathon_service_config.get_min_instances(),
max_instances=marathon_service_config.get_max_instances(),
current_instances=current_instances,
zookeeper_path=zookeeper_path,
**autoscaling_params)
new_instance_count = marathon_service_config.limit_instance_count(
current_instances + autoscaling_amount)
if new_instance_count != current_instances:
write_to_log(
config=marathon_service_config,
line='Scaling from %d to %d instances (%s)' %
(current_instances, new_instance_count, humanize_error(error)),
)
set_instances_for_marathon_service(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
instance_count=new_instance_count,
)
else:
write_to_log(
config=marathon_service_config,
line='Staying at %d instances (%s)' %
(current_instances, humanize_error(error)),
level='debug',
)
def autoscale_marathon_instance(marathon_service_config, marathon_tasks, mesos_tasks):
current_instances = marathon_service_config.get_instances()
if len(marathon_tasks) != current_instances:
write_to_log(config=marathon_service_config,
line='Delaying scaling as marathon is either waiting for resources or is delayed')
return
autoscaling_params = marathon_service_config.get_autoscaling_params()
autoscaling_metrics_provider = get_service_metrics_provider(autoscaling_params.pop(SERVICE_METRICS_PROVIDER_KEY))
autoscaling_decision_policy = get_decision_policy(autoscaling_params.pop(DECISION_POLICY_KEY))
utilization = autoscaling_metrics_provider(marathon_service_config, marathon_tasks,
mesos_tasks, **autoscaling_params)
error = get_error_from_utilization(
utilization=utilization,
setpoint=autoscaling_params.pop('setpoint'),
current_instances=current_instances,
)
zookeeper_path = compose_autoscaling_zookeeper_root(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
)
autoscaling_amount = autoscaling_decision_policy(
error=error,
min_instances=marathon_service_config.get_min_instances(),
max_instances=marathon_service_config.get_max_instances(),
current_instances=current_instances,
zookeeper_path=zookeeper_path,
**autoscaling_params
)
new_instance_count = marathon_service_config.limit_instance_count(current_instances + autoscaling_amount)
if new_instance_count != current_instances:
write_to_log(
config=marathon_service_config,
line='Scaling from %d to %d instances (%s)' % (
current_instances, new_instance_count, humanize_error(error)),
)
set_instances_for_marathon_service(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
instance_count=new_instance_count,
)
else:
write_to_log(
config=marathon_service_config,
line='Staying at %d instances (%s)' % (current_instances, humanize_error(error)),
level='debug',
)
def update_autoscaler_count(request):
service = request.swagger_data.get("service")
instance = request.swagger_data.get("instance")
desired_instances = request.swagger_data.get(
"json_body")["desired_instances"]
if not isinstance(desired_instances, int):
error_message = 'The provided body does not have an integer value for "desired_instances": {}'.format(
request.swagger_data.get("json_body"))
raise ApiFailure(error_message, 500)
try:
service_config = load_marathon_service_config(
service=service,
instance=instance,
cluster=settings.cluster,
soa_dir=settings.soa_dir,
load_deployments=False,
)
except Exception:
error_message = f"Unable to load service config for {service}.{instance}"
raise ApiFailure(error_message, 404)
max_instances = service_config.get_max_instances()
if max_instances is None:
error_message = f"Autoscaling is not enabled for {service}.{instance}"
raise ApiFailure(error_message, 404)
min_instances = service_config.get_min_instances()
# Dump whatever number from the client to zk. get_instances() will limit
# readings from zk to [min_instances, max_instances].
set_instances_for_marathon_service(service=service,
instance=instance,
instance_count=desired_instances)
status = "SUCCESS"
if desired_instances > max_instances:
desired_instances = max_instances
status = (
"WARNING desired_instances is greater than max_instances %d" %
max_instances)
elif desired_instances < min_instances:
desired_instances = min_instances
status = ("WARNING desired_instances is less than min_instances %d" %
min_instances)
response_body = {"desired_instances": desired_instances, "status": status}
return Response(json_body=response_body, status_code=202)
def update_autoscaler_count(request):
service = request.swagger_data.get('service')
instance = request.swagger_data.get('instance')
desired_instances = request.swagger_data.get(
'json_body')['desired_instances']
try:
service_config = load_marathon_service_config(
service=service,
instance=instance,
cluster=settings.cluster,
soa_dir=settings.soa_dir,
load_deployments=False,
)
except Exception:
error_message = 'Unable to load service config for %s.%s' % (service,
instance)
raise ApiFailure(error_message, 404)
max_instances = service_config.get_max_instances()
if max_instances is None:
error_message = 'Autoscaling is not enabled for %s.%s' % (service,
instance)
raise ApiFailure(error_message, 404)
min_instances = service_config.get_min_instances()
# Dump whatever number from the client to zk. get_instances() will limit
# readings from zk to [min_instances, max_instances].
set_instances_for_marathon_service(service=service,
instance=instance,
instance_count=desired_instances)
status = 'SUCCESS'
if desired_instances > max_instances:
desired_instances = max_instances
status = 'WARNING desired_instances is greater than max_instances %d' % max_instances
elif desired_instances < min_instances:
desired_instances = min_instances
status = 'WARNING desired_instances is less than min_instances %d' % min_instances
response_body = {'desired_instances': desired_instances, 'status': status}
return Response(json_body=response_body, status_code=202)
def update_autoscaler_count(request):
service = request.swagger_data.get('service')
instance = request.swagger_data.get('instance')
desired_instances = request.swagger_data.get('json_body')['desired_instances']
try:
service_config = load_marathon_service_config(
service=service,
instance=instance,
cluster=settings.cluster,
soa_dir=settings.soa_dir,
load_deployments=False,
)
except Exception:
error_message = 'Unable to load service config for %s.%s' % (service, instance)
raise ApiFailure(error_message, 404)
max_instances = service_config.get_max_instances()
if max_instances is None:
error_message = 'Autoscaling is not enabled for %s.%s' % (service, instance)
raise ApiFailure(error_message, 404)
min_instances = service_config.get_min_instances()
# Dump whatever number from the client to zk. get_instances() will limit
# readings from zk to [min_instances, max_instances].
set_instances_for_marathon_service(service=service, instance=instance, instance_count=desired_instances)
status = 'SUCCESS'
if desired_instances > max_instances:
desired_instances = max_instances
status = 'WARNING desired_instances is greater than max_instances %d' % max_instances
elif desired_instances < min_instances:
desired_instances = min_instances
status = 'WARNING desired_instances is less than min_instances %d' % min_instances
response_body = {'desired_instances': desired_instances, 'status': status}
return Response(json_body=response_body, status_code=202)
def autoscale_marathon_instance(marathon_service_config, system_paasta_config,
marathon_tasks, mesos_tasks):
current_instances = marathon_service_config.get_instances()
task_data_insufficient = is_task_data_insufficient(marathon_service_config,
marathon_tasks,
current_instances)
autoscaling_params = marathon_service_config.get_autoscaling_params()
log_utilization_data = {}
utilization = get_utilization(
marathon_service_config=marathon_service_config,
system_paasta_config=system_paasta_config,
autoscaling_params=autoscaling_params,
log_utilization_data=log_utilization_data,
marathon_tasks=marathon_tasks,
mesos_tasks=mesos_tasks,
)
error = get_error_from_utilization(
utilization=utilization,
setpoint=autoscaling_params['setpoint'],
current_instances=current_instances,
)
new_instance_count = get_new_instance_count(
utilization=utilization,
error=error,
autoscaling_params=autoscaling_params,
current_instances=current_instances,
marathon_service_config=marathon_service_config,
num_healthy_instances=len(marathon_tasks),
)
safe_downscaling_threshold = int(current_instances * 0.7)
if new_instance_count != current_instances:
if new_instance_count < current_instances and task_data_insufficient:
write_to_log(
config=marathon_service_config,
line=
'Delaying scaling *down* as we found too few healthy tasks running in marathon. '
'This can happen because tasks are delayed/waiting/unhealthy or because we are '
'waiting for tasks to be killed. Will wait for sufficient healthy tasks before '
'we make a decision to scale down.',
)
return
if new_instance_count == safe_downscaling_threshold:
write_to_log(
config=marathon_service_config,
line='Autoscaler clamped: %s' % str(log_utilization_data),
level='debug',
)
write_to_log(
config=marathon_service_config,
line='Scaling from %d to %d instances (%s)' % (
current_instances,
new_instance_count,
humanize_error(error),
),
)
set_instances_for_marathon_service(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
instance_count=new_instance_count,
)
else:
write_to_log(
config=marathon_service_config,
line='Staying at %d instances (%s)' %
(current_instances, humanize_error(error)),
level='debug',
)
meteorite_dims = {
'service_name': marathon_service_config.service,
'decision_policy': autoscaling_params[DECISION_POLICY_KEY],
'paasta_cluster': marathon_service_config.cluster,
'instance_name': marathon_service_config.instance,
}
if yelp_meteorite:
gauge = yelp_meteorite.create_gauge('paasta.service.instances',
meteorite_dims)
gauge.set(new_instance_count)
gauge = yelp_meteorite.create_gauge('paasta.service.max_instances',
meteorite_dims)
gauge.set(marathon_service_config.get_max_instances())
gauge = yelp_meteorite.create_gauge('paasta.service.min_instances',
meteorite_dims)
gauge.set(marathon_service_config.get_min_instances())
def autoscale_marathon_instance(marathon_service_config, marathon_tasks,
mesos_tasks):
current_instances = marathon_service_config.get_instances()
too_many_instances_running = len(marathon_tasks) > int(
(1 + MAX_TASK_DELTA) * current_instances)
too_few_instances_running = len(marathon_tasks) < int(
(1 - MAX_TASK_DELTA) * current_instances)
if too_many_instances_running or too_few_instances_running:
if current_instances < marathon_service_config.get_min_instances():
write_to_log(
config=marathon_service_config,
line=
'Scaling from %d to %d instances because we are below min_instances'
% (current_instances,
marathon_service_config.get_min_instances()))
set_instances_for_marathon_service(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
instance_count=marathon_service_config.get_min_instances())
else:
write_to_log(
config=marathon_service_config,
line=
'Delaying scaling as we found too many or too few tasks running in marathon. '
'This can happen because tasks are delayed/waiting/unhealthy or because we are '
'waiting for tasks to be killed.')
return
autoscaling_params = marathon_service_config.get_autoscaling_params()
autoscaling_metrics_provider = get_service_metrics_provider(
autoscaling_params.pop(SERVICE_METRICS_PROVIDER_KEY))
autoscaling_decision_policy = get_decision_policy(
autoscaling_params.pop(DECISION_POLICY_KEY))
log_utilization_data = {}
utilization = autoscaling_metrics_provider(
marathon_service_config=marathon_service_config,
marathon_tasks=marathon_tasks,
mesos_tasks=mesos_tasks,
log_utilization_data=log_utilization_data,
**autoscaling_params)
error = get_error_from_utilization(
utilization=utilization,
setpoint=autoscaling_params.pop('setpoint'),
current_instances=current_instances,
)
zookeeper_path = compose_autoscaling_zookeeper_root(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
)
autoscaling_amount = autoscaling_decision_policy(
error=error,
min_instances=marathon_service_config.get_min_instances(),
max_instances=marathon_service_config.get_max_instances(),
current_instances=current_instances,
zookeeper_path=zookeeper_path,
**autoscaling_params)
# Limit downscaling by 30% of current_instances until we find out what is
# going on in such situations
safe_downscaling_threshold = int(current_instances * 0.7)
new_instance_count = max(current_instances + autoscaling_amount,
safe_downscaling_threshold)
new_instance_count = marathon_service_config.limit_instance_count(
new_instance_count)
if new_instance_count != current_instances:
if new_instance_count == safe_downscaling_threshold:
write_to_log(
config=marathon_service_config,
line='Autoscaler clamped: %s' % str(log_utilization_data),
level='debug',
)
write_to_log(
config=marathon_service_config,
line='Scaling from %d to %d instances (%s)' %
(current_instances, new_instance_count, humanize_error(error)),
)
set_instances_for_marathon_service(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
instance_count=new_instance_count,
)
else:
write_to_log(
config=marathon_service_config,
line='Staying at %d instances (%s)' %
(current_instances, humanize_error(error)),
level='debug',
)
def autoscale_marathon_instance(marathon_service_config, marathon_tasks,
mesos_tasks):
current_instances = marathon_service_config.get_instances()
task_data_insufficient = is_task_data_insufficient(marathon_service_config,
marathon_tasks,
current_instances)
autoscaling_params = marathon_service_config.get_autoscaling_params()
log_utilization_data = {}
utilization = get_utilization(
marathon_service_config=marathon_service_config,
autoscaling_params=autoscaling_params,
log_utilization_data=log_utilization_data,
marathon_tasks=marathon_tasks,
mesos_tasks=mesos_tasks)
error = get_error_from_utilization(
utilization=utilization,
setpoint=autoscaling_params['setpoint'],
current_instances=current_instances,
)
new_instance_count = get_new_instance_count(
utilization=utilization,
error=error,
autoscaling_params=autoscaling_params,
current_instances=current_instances,
marathon_service_config=marathon_service_config,
num_healthy_instances=len(marathon_tasks),
)
safe_downscaling_threshold = int(current_instances * 0.7)
if new_instance_count != current_instances:
if new_instance_count < current_instances and task_data_insufficient:
write_to_log(
config=marathon_service_config,
line=
'Delaying scaling *down* as we found too few healthy tasks running in marathon. '
'This can happen because tasks are delayed/waiting/unhealthy or because we are '
'waiting for tasks to be killed. Will wait for sufficient healthy tasks before '
'we make a decision to scale down.')
return
if new_instance_count == safe_downscaling_threshold:
write_to_log(
config=marathon_service_config,
line='Autoscaler clamped: %s' % str(log_utilization_data),
level='debug',
)
write_to_log(
config=marathon_service_config,
line='Scaling from %d to %d instances (%s)' %
(current_instances, new_instance_count, humanize_error(error)),
)
set_instances_for_marathon_service(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
instance_count=new_instance_count,
)
else:
write_to_log(
config=marathon_service_config,
line='Staying at %d instances (%s)' %
(current_instances, humanize_error(error)),
level='debug',
)
def autoscale_marathon_instance(
marathon_service_config: MarathonServiceConfig,
system_paasta_config: SystemPaastaConfig,
marathon_tasks: Sequence[MarathonTask],
mesos_tasks: Sequence[Task],
) -> None:
try:
with create_autoscaling_lock(marathon_service_config.service,
marathon_service_config.instance):
current_instances = marathon_service_config.get_instances()
task_data_insufficient = is_task_data_insufficient(
marathon_service_config=marathon_service_config,
marathon_tasks=marathon_tasks,
current_instances=current_instances,
)
autoscaling_params = marathon_service_config.get_autoscaling_params(
)
log_utilization_data: Mapping = {}
utilization = get_utilization(
marathon_service_config=marathon_service_config,
system_paasta_config=system_paasta_config,
autoscaling_params=autoscaling_params,
log_utilization_data=log_utilization_data,
marathon_tasks=marathon_tasks,
mesos_tasks=mesos_tasks,
)
error = get_error_from_utilization(
utilization=utilization,
setpoint=autoscaling_params["setpoint"],
current_instances=current_instances,
)
num_healthy_instances = len(marathon_tasks)
new_instance_count = get_new_instance_count(
utilization=utilization,
error=error,
autoscaling_params=autoscaling_params,
current_instances=current_instances,
marathon_service_config=marathon_service_config,
num_healthy_instances=num_healthy_instances,
persist_data=(not task_data_insufficient),
)
safe_downscaling_threshold = int(current_instances * 0.7)
_record_autoscaling_decision(
marathon_service_config=marathon_service_config,
autoscaling_params=autoscaling_params,
utilization=utilization,
log_utilization_data=log_utilization_data,
error=error,
current_instances=current_instances,
num_healthy_instances=num_healthy_instances,
new_instance_count=new_instance_count,
safe_downscaling_threshold=safe_downscaling_threshold,
task_data_insufficient=task_data_insufficient,
)
if new_instance_count != current_instances:
if new_instance_count < current_instances and task_data_insufficient:
write_to_log(
config=marathon_service_config,
line=
"Delaying scaling *down* as we found too few healthy tasks running in marathon. "
"This can happen because tasks are delayed/waiting/unhealthy or because we are "
"waiting for tasks to be killed. Will wait for sufficient healthy tasks before "
"we make a decision to scale down.",
level="debug",
)
return
else:
set_instances_for_marathon_service(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
instance_count=new_instance_count,
)
write_to_log(
config=marathon_service_config,
line="Scaling from %d to %d instances (%s)" % (
current_instances,
new_instance_count,
humanize_error(error),
),
level="event",
)
except LockHeldException:
log.warning(
"Skipping autoscaling run for {service}.{instance} because the lock is held"
.format(
service=marathon_service_config.service,
instance=marathon_service_config.instance,
))
