Load Scheduling with Average Latency Feedback

Introduction

This policy detects traffic overloads and cascading failure build-up by comparing the real-time latency with a historical average. A gradient controller calculates a proportional response to limit the accepted token (or request) rate. The token rate is reduced by a multiplicative factor when the service is overloaded, and increased by an additive factor while the service is no longer overloaded.

At a high level, this policy works as follows:

• Latency trend-based overload detection: A Flux Meter is used to gather latency metrics from a service control point. The historical latency over a large time window (30 minutes by default) is used to establish a long-term trend that can be compared to the current latency to detect overloads.
• Gradient Controller: Set point latency and current latency signals are fed to the gradient controller that calculates the proportional response to adjust the accepted token rate (Control Variable).
• Integral Optimizer: When the service is detected to be in the normal state, an integral optimizer is used to additively increase the accepted token rate of the service in each execution cycle of the circuit. This measured approach prevents accepting all the traffic at once after an overload, which can again lead to an overload.
• Load Scheduler: The accepted token rate at the service is throttled by a weighted-fair queuing scheduler. The output of the adjustments to accepted token rate made by gradient controller and optimizer logic are translated to a load multiplier that is synchronized with Aperture Agents through etcd. The load multiplier adjusts (increases or decreases) the token bucket fill rates based on the incoming token rate observed at each agent.

The following PromQL query (with appropriate filters) is used as SIGNAL for the load scheduler:

sum(increase(flux_meter_sum))/sum(increase(flux_meter_count))

info

See reference for the AIMDLoadScheduler component that is used within this blueprint.

info

See the use-case Adaptive Service Protection with Average Latency Feedback to see this blueprint in use.

Configuration

Blueprint name: load-scheduling/average-latency

Parameters

policy

Parameter	policy.components
Description	List of additional circuit components.
Type	Array of Object (aperture.spec.v1.Component)
Default Value	Expand []

Parameter	policy.policy_name
Description	Name of the policy.
Type	string
Default Value	__REQUIRED_FIELD__

Parameter	policy.resources
Description	Additional resources.
Type	Object (aperture.spec.v1.Resources)
Default Value	Expand flow_control: classifiers: []

policy.load_scheduling_core

Parameter	policy.load_scheduling_core.dry_run
Description	Default configuration for setting dry run mode on Load Scheduler. In dry run mode, the Load Scheduler acts as a passthrough and does not throttle flows. This config can be updated at runtime without restarting the policy.
Type	Boolean
Default Value	false

Parameter	policy.load_scheduling_core.kubelet_overload_confirmations
Description	Overload confirmation signals from kubelet.
Type	Object (kubelet_overload_confirmations)
Default Value	Expand {}

Parameter	policy.load_scheduling_core.overload_confirmations
Description	List of overload confirmation criteria. Load scheduler can throttle flows when all of the specified overload confirmation criteria are met.
Type	Array of Object (overload_confirmation)
Default Value	Expand []

Parameter	policy.load_scheduling_core.aimd_load_scheduler
Description	Parameters for AIMD throttling strategy.
Type	Object (aperture.spec.v1.AIMDLoadSchedulerParameters)
Default Value	Expand alerter: alert_name: AIMD Load Throttling Event gradient: max_gradient: 1 min_gradient: 0.1 slope: -1 load_multiplier_linear_increment: 0.025 load_scheduler: selectors: - control_point: __REQUIRED_FIELD__ max_load_multiplier: 2

policy.latency_baseliner

Parameter	policy.latency_baseliner.flux_meter
Description	Flux Meter defines the scope of latency measurements.
Type	Object (aperture.spec.v1.FluxMeter)
Default Value	Expand selectors: - control_point: __REQUIRED_FIELD__

Parameter	policy.latency_baseliner.latency_tolerance_multiplier
Description	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	Number (double)
Default Value	1.25

Parameter	policy.latency_baseliner.long_term_query_interval
Description	Interval for long-term latency query, i.e., how far back in time the query is run. The value should be a string representing the duration in seconds.
Type	string
Default Value	1800s

Parameter	policy.latency_baseliner.long_term_query_periodic_interval
Description	Periodic interval for long-term latency query, i.e., how often the query is run. The value should be a string representing the duration in seconds.
Type	string
Default Value	30s

---

Schemas

driver_criteria

Parameter	enabled
Description	Enables the driver.
Type	Boolean
Default Value	__REQUIRED_FIELD__

Parameter	threshold
Description	Threshold for the driver.
Type	Number (double)
Default Value	__REQUIRED_FIELD__

---

overload_confirmation_driver

Parameter	pod_cpu
Description	The driver for using CPU usage as overload confirmation.
Type	Object (driver_criteria)
Default Value	Expand {}

Parameter	pod_memory
Description	The driver for using memory usage as overload confirmation.
Type	Object (driver_criteria)
Default Value	Expand {}

---

kubelet_overload_confirmations

Parameter	criteria
Description	Criteria for overload confirmation.
Type	Object (overload_confirmation_driver)
Default Value	__REQUIRED_FIELD__

Parameter	infra_context
Description	Kubernetes selector for scraping metrics.
Type	Object (aperture.spec.v1.KubernetesObjectSelector)
Default Value	__REQUIRED_FIELD__

---

overload_confirmation

Parameter	operator
Description	The operator for the overload confirmation criteria. oneof: `gt | lt | gte | lte | eq | neq`
Type	string
Default Value

Parameter	query_string
Description	The Prometheus query to be run. Must return a scalar or a vector with a single element.
Type	string
Default Value

Parameter	threshold
Description	The threshold for the overload confirmation criteria.
Type	Number (double)
Default Value

---

Dynamic Configuration

note

The following configuration parameters can be dynamically configured at runtime, without reloading the policy.

Parameters

Parameter	dry_run
Description	Dynamic configuration for setting dry run mode at runtime without restarting this policy. In dry run mode the scheduler acts as pass through to all flow and does not queue flows. It is useful for observing the behavior of load scheduler without disrupting any real traffic.
Type	Boolean
Default Value	__REQUIRED_FIELD__

---