To clarify the 74% number - that’s relative to no hedging (or static hedging) in a straggler-heavy setup.
The interesting part here is less the absolute gain and more that adaptive hedging matches the best hand-tuned static threshold without requiring manual tuning, even as latency distributions shift.
In real systems where latency isn’t stationary, that’s where this approach should have more advantage.
One thing I found interesting is that even same-target hedging helps behind load balancers, since you’re effectively racing different backend replicas.
But I’m still thinking through cases where this could backfire (overload amplification, etc.)
The interesting part here is less the absolute gain and more that adaptive hedging matches the best hand-tuned static threshold without requiring manual tuning, even as latency distributions shift.
In real systems where latency isn’t stationary, that’s where this approach should have more advantage.
reply