Tag
3 entries
We present AdaptRaft, a variant of the Raft consensus algorithm that dynamically adjusts its timing parameters using a lightweight online learning model, reducing tail latency by up to 34% in geo-distributed deployments.
A modified Raft implementation with an online bandit learner that adapts election and heartbeat timeouts to observed network conditions. Companion codebase for the SOSP 2024 paper.
We explore the applicability of learned index structures to distributed key-value stores, demonstrating that model-based indices can reduce lookup latency by 22% with a modest increase in memory footprint on skewed workloads.