Hedged Request
also called Request Hedging, Tied Request
Sending a duplicate of a request to a second replica after a short delay and using whichever response returns first, to cut tail latency.
The observation behind it, from Google's "The Tail at Scale", is that tail latency is usually caused by transient conditions on a particular server — a garbage collection pause, a cold cache, a contended lock, a noisy neighbour — not by the request being inherently slow. So the same request sent elsewhere will very likely be fast.
The mechanism: send the request, and if no response has arrived by roughly the p95 latency, send a second copy to another replica and take whichever answers first. Because only around 5% of requests trigger a hedge, the extra load is small while the improvement at p99 can be large.
Conditions for using it: the operation must be idempotent or the duplicate must be safely cancellable, there must be spare capacity (hedging into a saturated system makes things worse), and the trigger delay must be tuned from the actual latency distribution rather than guessed.
The refinement is a tied request, where the replicas know about each other and the loser cancels its work — which removes most of the wasted capacity.