Every call to another service can fail, and the way a client retries decides whether a short blip stays short. Retrying at once and in lockstep turns one slow second into a flood of traffic; retrying past the caller's own timeout wastes work nobody will read.
These tasks build the standard defenses as pure functions, so they can be tested without clocks or networks. Exponential backoff with full jitter spreads clients out. A time budget stops retries that can no longer help. A circuit breaker stops calling a service that is clearly down and sends one trial request to notice recovery. And a retry is only allowed when the failure might pass and repeating the request can't do harm, which for a POST means an idempotency key.
Random numbers and times are passed in as data, which is how such code is tested in production projects too.
Out: [0.5, 1.0, 2.0, 4.0, 8.0, 10.0]. The cap keeps late retries reasonable.