adaptive-replica-selection · research

Adaptive replica selection

Adaptive replica selection routes a request among eligible data replicas using changing performance or load information rather than a fixed choice.

C3 combines response-carried feedback with replica ranking, distributed rate control and backpressure, without request reissues. Feedback is delayed, so the design accounts for concurrent outstanding requests and the risk of clients concentrating on one replica. [1]

The paper evaluates a Cassandra prototype using synthetic YCSB workloads on EC2. It does not establish a universal speedup for other engines or current deployments. [1]

Relations

related

hedged-request, tied-request, tail-latency, asynchronous-replication

Citations

1. Suresh, Canini, Schmid and Feldmann. C3: Cutting Tail Latency in Cloud Data Stores via Adaptive Replica Selection. NSDI 2015, sections 3-5. (accessed 2026-10-17)

Further reading

2. Hesela. Hedged reads: replica dependence and cancellation work, with exact synthetic examples. (accessed 2026-10-17)

JSON: /data/terms/adaptive-replica-selection.json