Different retry limits help protect important service requests

When independently managed services become miscoordinated, an overloaded downstream service can reject valuable requests. Retries can make that overload worse by adding traffic. Can retry rules give important requests a better chance without requiring services to coordinate or sense overload?

DiffTry assigns each request class a local retry limit, using the limits to shift overload rejections toward lower-priority traffic.

The paper asks whether retries can protect valuable requests when a downstream microservice is overloaded, without waiting for services to coordinate or detect overload.

DiffTry assigns request classes different retry limits locally, so requests with fewer permitted retries are more likely to be dropped when service remains unavailable.

In a modeled hotel-reservation network under the heaviest tested flood, high-priority rejection fell from about 54% with equal retry limits to about 15% with unequal limits, while low-priority rejection rose to about 61%.

A live deployment's measured rejection was close to the model's prediction, supporting the claim that local retry choices can produce network-wide prioritization in this tested setup.

The general question of whether the network equations have a unique solution for cyclic service networks remains open.

The live comparison also uses the application's native fan-out, while the analytical model assumes probabilistic routing, so the close match applies to this tested application and setup.