Hedged Request
Send a duplicate request to a second replica if the first is slow, and take whichever returns first — to cut tail latency.
A hedged request attacks tail latency: if a request to one replica has not responded within, say, the 95th-percentile time, the client sends a second copy to another replica and uses whichever answers first, cancelling the loser. It trades a little extra load for dramatically better p99.
Worked example: a read that usually takes 10ms but occasionally stalls at 200ms (a slow node, a GC pause) gets hedged after 20ms to a second replica that returns in 10ms — the user sees about 30ms instead of 200ms. Gotcha: hedging only helps when slowness is per-node and random, not a global overload — hedging during a real overload just adds load and makes things worse — so it must be capped (only hedge a small fraction of requests) and paired with the ability to cancel the redundant call.