Auto Scaler

info

See also Auto Scaler reference

Auto Scaler is a high-level component in Aperture that performs auto-scaling. It can interface with infrastructure APIs such as Kubernetes to automatically adjust the number of instances or resources allocated to a service to meet changing workload demands. Auto Scaler is designed to ensure that the service is scaled out to meet demand and scaled in when demand is low. Scaling out is done more aggressively than scaling in to ensure optimal performance and availability.

• Controllers: Auto Scaler leverages controllers (for example, Gradient Controller) to make scaling decisions. A Controller can be configured for either scaling in or out, and defines the criteria that determine when to scale. Controllers process one or more input signals to compute a desired scale value. By configuring Controllers, you can fine-tune the auto-scaling behavior to meet the specific needs of your service. See Gradient Controller for more details.
• A scale-in Controller is active only when its output is smaller than the actual scale value. A scale-out Controller is active only when its output is larger than the actual scale value. For example, the actual number of replicas of a Kubernetes Deployment. An inactive Controller does not contribute to the scaling decision.
• Scale decisions from multiple active Controllers are combined by the Auto Scaler by taking the largest scale value.
• Maximum scale-in and scale-out step sizes: The amount of scaling that happens at a time is limited by the maximum scale-in and scale-out step sizes. This is to prevent large-scale changes from happening at once.
• Cooldown periods: There are cooldown periods defined individually for scale-out and scale-in. The Auto Scaler won't scale-out or scale-in again until the cooldown period has elapsed. The intention of cooldowns is to make the changes gradually and observe their effect to prevent overdoing either scale-in or scale-out.
  • Scale-in cooldown: The Auto Scaler won't scale-in again until the cooldown period has elapsed. If there is a scale-out decision, the scale-in cooldown is reset. This immediately corrects excessive scale-in.
  • Scale-out cooldown: The Auto Scaler won't scale-out again until the cooldown period has elapsed. If there is a scale-out decision which is much larger than the current scale value, the scale-out cooldown is reset. This is done to accommodate any urgent need for scale-out.

Gradient Controller

The Gradient Controller computes a desired scale value based on a signal and setpoint. The gradient controller tries to adjust the scale value proportionally to the relative difference between setpoint and signal.

The gradient describes a corrective factor that should be applied to the scale value to get the signal closer to the setpoint. It's computed as follows:

$$\text{gradient} = \left(\frac{\text{signal}}{\text{setpoint}}\right)^{\text{slope}}$$

gradient is then clamped to [1.0, max_gradient] range for the scale-out controller and [min_gradient, 1.0] range for the scale-in controller.

The output of the gradient controller is computed as follows:

$$\text{desired\_scale} = \text{gradient}_{\text{clamped}} \cdot \text{actual\_scale}.$$