DocsEmissary-ingressCircuit breakers
Circuit breakers
Circuit breakers are a powerful technique to improve resilience. By preventing additional connections or requests to an overloaded service, circuit breakers limit the "blast radius" of an overloaded service. By design, Emissary-ingress circuit breakers are distributed, i.e., different Emissary-ingress instances do not coordinate circuit breaker information.
Circuit breaker configuration
A default circuit breaking configuration can be set for all
Emissary-ingress resources in the ambassador
Module
, or set to a different value on a
per-resource basis for Mappings
,
TCPMappings
, and
AuthServices
.
The circuit_breakers
attribute configures circuit breaking. The following fields are supported:
Key | Default value | Description |
---|---|---|
priority | default | Specifies the priority to which the circuit breaker settings apply to; it can be set to either default or high . |
max_connections | 1024 | Specifies the maximum number of connections that Emissary-ingress will make to the services. In practice, this is more applicable to HTTP/1.1 than HTTP/2. |
max_pending_requests | 1024 | Specifies the maximum number of requests that will be queued while waiting for a connection. In practice, this is more applicable to HTTP/1.1 than HTTP/2. |
max_requests | 1024 | Specifies the maximum number of parallel outstanding requests to an upstream service. In practice, this is more applicable to HTTP/2 than HTTP/1.1. |
max_retries | 3 | Specifies the maximum number of parallel retries allowed to an upstream service. |
The max_*
fields can be reduced to shrink the "blast radius," or
increased to enable Emissary-ingress to handle a larger number of
concurrent requests.
Examples
Circuit breakers defined on a single Mapping
:
Circuit breakers defined on a single AuthService
:
A global circuit breaker:
Circuit breakers and automatic retries
Circuit breakers are best used in conjunction with automatic retries. Here are some examples:
- You've configured automatic retries for failed requests to a service. Your service is under heavy load, and starting to time out on servicing requests. In this case, automatic retries can exacerbate your problem, increasing the total request volume by 2x or more. By aggressively circuit breaking, you can mitigate failure in this scenario.
- To circuit break when services are slow, you can combine circuit breakers with retries. Reduce the time out for retries, and then set a circuit breaker that detects many retries. In this setup, if your service doesn't respond quickly, a flood of retries will occur, which can then trip the circuit breaker.
Note that setting circuit breaker thresholds requires careful monitoring and experimentation. We recommend you start with conservative values for circuit breakers and adjust them over time.
More about circuit breakers
Responses from a broken circuit contain the x-envoy-overloaded
header.
The following are the default values for circuit breaking if nothing is specified:
Circuit breaker metrics are exposed in StatsD. For more information about the specific statistics, see the Envoy documentation.