The overload manager is an extensible component for protecting the Envoy server from overload with respect to various system resources (such as memory, cpu or file descriptors) due to too many client connections or requests. This is distinct from circuit breaking which is primarily aimed at protecting upstream services.
The overload manager is configured by specifying a set of resources to monitor and a set of overload actions that will be taken when some of those resources exceed certain pressure thresholds.
The overload manager works by periodically polling the pressure of a set of resources, feeding those through triggers, and taking actions based on the triggers. The set of resource monitors, triggers, and actions are specified at startup.
A resource is a thing that can be monitored by the overload manager, and whose pressure is represented by a real value in the range [0, 1]. The pressure of a resource is evaluated by a resource monitor. See the configuration page for setting up resource monitors.
Triggers are evaluated on each resource pressure update, and convert a resource pressure value into an action state. An action state has a value in the range [0, 1], and is categorized into one of two groups:
the resource pressure is below the configured saturation point; action may be taken
the resource pressure is at or above the configured saturation point; drastic action should be taken
When a resource pressure value is updated, the relevant triggers are reevaluated. For each action with at least one trigger, the resulting action state is the maximum value over the configured triggers. What effect the action state has depends on the action’s configuration and implementation.
When a trigger changes state, the value is sent to registered actions, which can then affect how connections and requests are processed. Each action interprets the input states differently, and some may ignore the scaling state altogether, taking effect only when saturated.