Glossary of TermsΒΆ

Retry Storm

A retry storm is an undesirable client/server failure mode where one or more peers become unhealthy, causing clients to retry a significant fraction of requests. This has the effect of multiplying the volume of traffic sent to the unhealthy peers, exacerbating the problem.

For instance, consider two services, A -> B where B becomes overloaded and starts to fail a large fraction of requests. If A is configured to greedily optimize its client success rate by retrying requests, the total number of requests to B will be substantially amplified. B is the victim of a retry storm.

Back pressure

Back pressure is a way of signaling that a service is overloaded to its clients. This allows for graceful degradation of a system if one or more components are not performing well.

Nack

A nack is a Negative ACKnowledgement. A service responds with a nack to signal that it is refusing to process the request. Generally, this indicates that the request is safe to retry, but Finagle also supports non-retryable nacks. Responding with a nack is one way a service may exert back pressure.