Skip to main content
Version: development

Average Latency Feedback

Overview​

The response times of a service start to deteriorate when the service's underlying concurrency limit is surpassed. Consequently, a degradation in response latency can serve as a reliable signal for identifying service overload. This policy is designed to detect overload situations based on latency deterioration. During overload, the request rate is throttled so that latency gets restored back to an acceptable range.

Configuration​

This policy is based on the Load Scheduling with Average Latency Feedback blueprint. The latency of requests processed by the cart-service.prod.svc.cluster.local service is monitored. By analyzing the variance between current latency and a historical latency baseline, the policy facilitates the identification of potential service overload. A deviation of 1.1 from the baseline is considered as a signal of service overload.

To mitigate service overload, the requests to the cart-service.prod.svc.cluster.local service are passed through a load scheduler. The load scheduler reduces the request rate in overload scenarios, temporarily placing excess requests in a queue.

As service latency improves, indicating a return to normal operational state, the request rate is incrementally increased until it matches the incoming request rate. This responsive mechanism helps ensure that service performance is optimized while mitigating the risk of service disruptions due to overload.

The below values.yaml file can be generated by following the steps in the Installation section.

# yaml-language-server: $schema=../../../../../../blueprints/load-scheduling/average-latency/gen/definitions.json
# Generated values file for load-scheduling/average-latency blueprint
# Documentation/Reference for objects and parameters can be found at:
# https://docs.fluxninja.com/reference/blueprints/load-scheduling/average-latency

policy:
  # Name of the policy.
  # Type: string
  # Required: True
  policy_name: basic-service-protection
  service_protection_core:
    adaptive_load_scheduler:
      load_scheduler:
        # The selectors determine the flows that are protected by this policy.
        # Type: []aperture.spec.v1.Selector
        # Required: True
        selectors:
          - control_point: ingress
            service: cart-service.prod.svc.cluster.local
  latency_baseliner:
    # Tolerance factor beyond which the service is considered to be in overloaded state. E.g. if the long-term average of latency is L and if the tolerance is T, then the service is considered to be in an overloaded state if the short-term average of latency is more than L*T.
    # Type: float64
    latency_tolerance_multiplier: 1.1
    # Flux Meter defines the scope of latency measurements.
    # Type: aperture.spec.v1.FluxMeter
    # Required: True
    flux_meter:
      selectors:
        - control_point: ingress
          service: cart-service.prod.svc.cluster.local

Generated Policy

apiVersion: fluxninja.com/v1alpha1
kind: Policy
metadata:
  labels:
    fluxninja.com/validate: "true"
  name: basic-service-protection
spec:
  circuit:
    components:
    - flow_control:
        adaptive_load_scheduler:
          dry_run: false
          dry_run_config_key: dry_run
          in_ports:
            overload_confirmation:
              constant_signal:
                value: 1
            setpoint:
              signal_name: SETPOINT
            signal:
              signal_name: SIGNAL
          out_ports:
            desired_load_multiplier:
              signal_name: DESIRED_LOAD_MULTIPLIER
            observed_load_multiplier:
              signal_name: OBSERVED_LOAD_MULTIPLIER
          parameters:
            alerter:
              alert_name: Load Throttling Event
            gradient:
              max_gradient: 1
              min_gradient: 0.1
              slope: -1
            load_multiplier_linear_increment: 0.025
            load_scheduler:
              selectors:
              - control_point: ingress
                service: cart-service.prod.svc.cluster.local
            max_load_multiplier: 2
    - query:
        promql:
          evaluation_interval: 10s
          out_ports:
            output:
              signal_name: SIGNAL
          query_string: sum(increase(flux_meter_sum{flow_status="OK", flux_meter_name="basic-service-protection",
            policy_name="basic-service-protection"}[30s]))/sum(increase(flux_meter_count{flow_status="OK",
            flux_meter_name="basic-service-protection", policy_name="basic-service-protection"}[30s]))
    - query:
        promql:
          evaluation_interval: 30s
          out_ports:
            output:
              signal_name: SIGNAL_LONG_TERM
          query_string: sum(increase(flux_meter_sum{flow_status="OK", flux_meter_name="basic-service-protection",
            policy_name="basic-service-protection"}[1800s]))/sum(increase(flux_meter_count{flow_status="OK",
            flux_meter_name="basic-service-protection", policy_name="basic-service-protection"}[1800s]))
    - arithmetic_combinator:
        in_ports:
          lhs:
            signal_name: SIGNAL_LONG_TERM
          rhs:
            constant_signal:
              value: 1.1
        operator: mul
        out_ports:
          output:
            signal_name: SETPOINT
    evaluation_interval: 10s
  resources:
    flow_control:
      classifiers: []
      flux_meters:
        basic-service-protection:
          selectors:
          - control_point: ingress
            service: cart-service.prod.svc.cluster.local
    infra_meters: {}

info

Circuit Diagram for this policy.

Installation​

Generate a values file specific to the policy. This can be achieved using the command provided below.

aperturectl blueprints values --name=load-scheduling/average-latency --version=main --output-file=values.yaml

Adjust the values to match the application requirements. Use the following command to generate the policy.

aperturectl blueprints generate --name=load-scheduling/average-latency --values-file=values.yaml --output-dir=policy-gen --version=main

Apply the policy using the aperturectl CLI or kubectl.

Pass the --kube flag with aperturectl to directly apply the generated policy on a Kubernetes cluster in the namespace where the Aperture Controller is installed.

aperturectl apply policy --file=policy-gen/policies/load-scheduling.yaml --kube

Policy in Action​

To see the policy in action, the traffic is generated such that it starts within the service's capacity and then goes beyond the capacity after some time. Such a traffic pattern is repeated periodically. The below dashboard demonstrates that when latency spikes due to high traffic at cart-service.prod.svc.cluster.local, the controller throttles the rate of requests admitted into the service. This approach helps protect the service from becoming unresponsive and maintains the current latency within the tolerance limit (1.1) of historical latency.

Basic Service Protection

Dry Run Mode​

You can run this policy in the Dry Run mode by setting the policy.service_protection_core.dry_run parameter to true. In the Dry Run mode, the policy does not throttle the request rate while still evaluating the decisions it would take in each cycle. This is useful for evaluating the policy without impacting the service.

note

The Dry Run mode can also be toggled dynamically at runtime, without reloading the policy.

Demo Video​

The below demo video shows the basic service protection and workload prioritization policy in action within Aperture Playground.

Demo Video