Overload manager

The overload manager is configured in the Bootstrap overload_manager field.

An example configuration of the overload manager is shown below. It shows a configuration to drain HTTP/X connections when heap memory usage reaches 92% (configured via envoy.overload_actions.disable_http_keepalive), to stop accepting requests when heap memory usage reaches 95% (configured via envoy.overload_actions.stop_accepting_requests) and to stop accepting new TCP connections when memory usage reaches 95% (configured via envoy.load_shed_points.tcp_listener_accept).

refresh_interval:
  seconds: 0
  nanos: 250000000
resource_monitors:
  - name: "envoy.resource_monitors.fixed_heap"
    typed_config:
      "@type": type.googleapis.com/envoy.extensions.resource_monitors.fixed_heap.v3.FixedHeapConfig
      max_heap_size_bytes: 2147483648
actions:
  - name: "envoy.overload_actions.disable_http_keepalive"
    triggers:
      - name: "envoy.resource_monitors.fixed_heap"
        threshold:
          value: 0.92
  - name: "envoy.overload_actions.stop_accepting_requests"
    triggers:
      - name: "envoy.resource_monitors.fixed_heap"
        threshold:
          value: 0.95
 loadshed_points:
   - name: "envoy.load_shed_points.tcp_listener_accept"
     triggers:
       - name: "envoy.resource_monitors.fixed_heap"
         threshold:
           value: 0.95

Resource monitors

The overload manager uses Envoy’s extension framework for defining resource monitors. Envoy’s builtin resource monitors are listed here.

Triggers

Triggers connect resource monitors to actions. There are two types of triggers supported:

Type	Description
threshold	Sets the action state to 1 (= saturated) when the resource pressure is above a threshold, and to 0 otherwise.
scaled	Sets the action state to 0 when the resource pressure is below the scaling_threshold, `(pressure - scaling_threshold)/(saturation_threshold - scaling_threshold)` when `scaling_threshold < pressure < saturation_threshold`, and to 1 (saturated) when the pressure is above the saturation_threshold.”

Overload actions

The following overload actions are supported:

Name	Description
envoy.overload_actions.stop_accepting_requests	Envoy will immediately respond with a 503 response code to new requests
envoy.overload_actions.disable_http_keepalive	Envoy will drain HTTP/2 and HTTP/3 connections using `GOAWAY` with a drain grace period. For HTTP/1, Envoy will set a drain timer to close the more idle recently used connections.
envoy.overload_actions.stop_accepting_connections	Envoy will stop accepting new network connections on its configured listeners
envoy.overload_actions.reject_incoming_connections	Envoy will reject incoming connections on its configured listeners without processing any data
envoy.overload_actions.shrink_heap	Envoy will periodically try to shrink the heap by releasing free memory to the system
envoy.overload_actions.reduce_timeouts	Envoy will reduce the waiting period for a configured set of timeouts. See below for details on configuration.
envoy.overload_actions.reset_high_memory_stream	Envoy will reset expensive streams to terminate them. See below for details on configuration.

Load Shed Points

Load Shed Points are similar to overload actions as they are dependent on a given trigger to activate which determines whether Envoy ends up shedding load at the given point in a connection or stream lifecycle.

For a given request on a newly created connection, we can think of the configured load shed points as a decision tree at key junctions of a connection / stream lifecycle. While a connection / stream might pass one junction, it is possible that later on the conditions might change causing Envoy to shed load at a later junction.

In comparision to analogous overload actions, Load Shed Points are more reactive to changing conditions, especially in cases of large traffic spikes. Overload actions can be better suited in cases where Envoy is deciding to shed load but the worker threads aren’t actively processing the connections or streams that Envoy wants to shed. For example envoy.overload_actions.reset_high_memory_stream can reset streams that are using a lot of memory even if those streams aren’t actively making progress.

Compared to overload actions, Load Shed Points are also more flexible to integrate custom (e.g. company inteneral) Load Shed Points as long as the extension has access to the Overload Manager to request the custom Load Shed Point.

The following core load shed points are supported:

Name	Description
envoy.load_shed_points.tcp_listener_accept	Envoy will reject (close) new TCP connections. This occurs before the Listener Filter Chain is created.
envoy.load_shed_points.http_connection_manager_decode_headers	Envoy will reject new HTTP streams by sending a local reply. This occurs right after the http codec has finished parsing headers but before the HTTP Filter Chain is instantiated.
envoy.load_shed_points.http1_server_abort_dispatch	Envoy will reject processing HTTP1 at the codec level. If a response has not yet started, Envoy will send a local reply. Envoy will then close the connection.
envoy.load_shed_points.http2_server_go_away_on_dispatch	Envoy will send a `GOAWAY` while processing HTTP2 requests at the codec level which will eventually drain the HTTP/2 connection.

Reducing timeouts

The envoy.overload_actions.reduce_timeouts overload action will reduce the amount of time Envoy will spend waiting for some interactions to finish in response to resource pressure. The amount of reduction can be configured per timeout type by specifying the minimum timer value to use when the triggering resource monitor detects saturation. The minimum value for each timeout can be specified either by providing a scale factor to apply to the configured maximum, or as a concrete duration value.

As an example, here is a single overload action entry that enables timeout reduction:

name: "envoy.overload_actions.reduce_timeouts"
triggers:
  - name: "envoy.resource_monitors.fixed_heap"
    scaled:
      scaling_threshold: 0.85
      saturation_threshold: 0.95
typed_config:
  "@type": type.googleapis.com/envoy.config.overload.v3.ScaleTimersOverloadActionConfig
  timer_scale_factors:
    - timer: HTTP_DOWNSTREAM_CONNECTION_IDLE
      min_timeout: 2s

It configures the overload manager to change the amount of time that HTTP connections are allowed to remain idle before being closed in response to heap size. When the heap usage is less than 85%, idle connections will time out at their usual time, which is configured through HttpConnectionManager.common_http_protocol_options.idle_timeout. When the heap usage is at or above 95%, idle connections will be closed after the specified min_timeout, here 2 seconds. If the heap usage is between 85% and 95%, the idle connection timeout will vary between those two based on the formula for the scaled trigger So if RouteAction.idle_timeout = 600 seconds and heap usage is at 92%, idle connections will time out after \(2s + (600s - 2s) \cdot (95\% - 92\%) / (95\% - 85\%) = 181.4s\).

Note in the example that the minimum idle time is specified as an absolute duration. If, instead, min_timeout: 2s were to be replaced with min_scale: { value: 10 }, the minimum timer value would be computed based on the maximum (specified elsewhere). So if idle_timeout is again 600 seconds, then the minimum timer value would be \(10\% \cdot 600s = 60s\).

Limiting Active Connections

Currently, the only supported way to limit the total number of active connections allowed across all listeners is via specifying an integer through the runtime key overload.global_downstream_max_connections. The connection limit is recommended to be less than half of the system’s file descriptor limit, to account for upstream connections, files, and other usage of file descriptors. If the value is unspecified, there is no global limit on the number of active downstream connections and Envoy will emit a warning indicating this at startup. To disable the warning without setting a limit on the number of active downstream connections, the runtime value may be set to a very large limit (~2e9). Listeners can opt out of this global connection limit by setting Listener.ignore_global_conn_limit to true. Similarly, you can opt out the admin listener by setting Admin.ignore_global_conn_limit. You may want to opt out a listener to be able to probe Envoy or collect stats while it is otherwise at its connection limit. Note that connections to listeners that opt out are still tracked and count towards the global limit.

If it is desired to only limit the number of downstream connections for a particular listener, per-listener limits can be set via the listener configuration.

One may simultaneously specify both per-listener and global downstream connection limits and the conditions will be enforced independently. For instance, if it is known that a particular listener should have a smaller number of open connections than others, one may specify a smaller connection limit for that specific listener and allow the global limit to enforce resource utilization among all listeners.

An example configuration can be found in the edge best practices document.

Reset Streams

Warning

Resetting streams via an overload action currently only works with HTTP2.

The envoy.overload_actions.reset_high_memory_stream overload action will reset expensive streams. This requires minimum_account_to_track_power_of_two to be configured via buffer_factory_config. To understand the memory class scheme in detail see minimum_account_to_track_power_of_two

As an example, here is a partial Overload Manager configuration with minimum threshold for tracking and a single overload action entry that resets streams:

buffer_factory_config:
  minimum_account_to_track_power_of_two: 20
actions:
  name: "envoy.overload_actions.reset_high_memory_stream"
  triggers:
    - name: "envoy.resource_monitors.fixed_heap"
      scaled:
        scaling_threshold: 0.85
        saturation_threshold: 0.95
...

We will only track streams using >= \(2^{minimum\_account\_to\_track\_power\_of\_two}\) worth of allocated memory in buffers. In this case, by setting the minimum_account_to_track_power_of_two to 20 we will track streams using >= 1MiB since \(2^{20}\) is 1MiB. Streams using >= 1MiB will be classified into 8 power of two sized buckets. Currently, the number of buckets is hardcoded to 8. For this example, the buckets are as follows:

Bucket index	Contains streams using
0	[1MiB,2MiB)
1	[2MiB,4MiB)
2	[4MiB,8MiB)
3	[8MiB,16MiB)
4	[16MiB,32MiB)
5	[32MiB,64MiB)
6	[64MiB,128MiB)
7	>= 128MiB

The above configuration also configures the overload manager to reset our tracked streams based on heap usage as a trigger. When the heap usage is less than 85%, no streams will be reset. When heap usage is at or above 85%, we start to reset buckets according to the strategy described below. When the heap usage is at 95% all streams using >= 1MiB memory are eligible for reset. This overload action will reset up to 50 streams (this is a hardcoded limit) per worker everytime the action is invoked. This is both to reduce the amount of streams that end up getting reset and to prevent the worker thread from locking up and triggering the Watchdog system.

Given that there are only 8 buckets, we partition the space with a gradation of \(gradation = (saturation\_threshold - scaling\_threshold)/8\). Hence at 85% heap usage we reset streams in the last bucket e.g. those using >= 128MiB. At \(85\% + 1 * gradation\) heap usage we reset streams in the last two buckets e.g. those using >= 64MiB, prioritizing the streams in the last bucket since there’s a hard limit on the number of streams we can reset per invokation. At \(85\% + 2 * gradation\) heap usage we reset streams in the last three buckets e.g. those using >= 32MiB. And so forth as the heap usage is higher.

It’s expected that the first few gradations shouldn’t trigger anything, unless there’s something seriously wrong e.g. in this example streams using >= 128MiB in buffers.

Statistics

Each configured resource monitor has a statistics tree rooted at overload.<name>. with the following statistics:

Name	Type	Description
pressure	Gauge	Resource pressure as a percent
failed_updates	Counter	Total failed attempts to update the resource pressure
skipped_updates	Counter	Total skipped attempts to update the resource pressure due to a pending update
refresh_interval_delay	Histogram	Latencies for the delay between overload manager resource refresh loops

Each configured overload action has a statistics tree rooted at overload.<name>. with the following statistics:

Name	Type	Description
active	Gauge	Active state of the action (0=scaling, 1=saturated)
scale_percent	Gauge	Scaled value of the action as a percent (0-99=scaling, 100=saturated)

Each configured Load Shed Point has a statistics tree rooted at overload.<name>. with the following statistics:

Name	Type	Description
scale_percent	Gauge	Scaled value of the action as a percent (0-99=scaling, 100=saturated)