Load Balancing

Linkerd uses a sophisticated EWMA load balancing algorithm which typically requires no configuration or tuning. However, it also supports special handling for backends which support rate limiting. This special handling can be enabled with Service annotations.

Load Biaser

When backends implement rate limiting and return HTTP 429 or gRPC RESOURCE_EXHAUSTED by default, the proxy treats these as successful responses from a load balancing perspective. Since these types of responses are typically very fast, Linkerd’s EWMA load balancing may actually send more traffic to these rate-limited endpoints. This can create a feedback loop where clients experience high 429 or RESOURCE_EXHAUSTED rates.

Linkerd can be configured to use a more sophisticated version of the EWMA load balancing algorithm which takes rate-limit responses (HTTP 429 or gRPC RESOURCE_EXHAUSTED) into account.

To enable Linkerd to use the Load Biaser for a Service, set the following annotation on the Service resource:

annotation	value
`balancer.alpha.linkerd.io/penalize-failures`	Enables the Load Biaser for this Service. Defaults to `false`.
`balancer.alpha.linkerd.io/load-biaser-penalty`	The latency value to inject for rate-limited responses and failures. Defaults to `5s`.
`balancer.alpha.linkerd.io/load-biaser-max-retry-after`	The maximum allowed value of a Retry-After header. Defaults to `300s`.