双利普希茨网络和区间锁增强的学习型索引OA

Learned indexes accelerate query processing by fitting the mapping between keys and positions,and have become a research hotspot.Mainstream methods employ piece-wise linear functions to improve fitting quality.However,when handling complex data distributions,they require a large number of segments,leading to index size inflation,fre-quent retraining,and degraded query performance.Some studies introduce distribution transformation to mitigate this issue,but existing methods suffer from high training and transformation overheads as well as limited stability.To address these challenges,this paper proposes BLELI(bi-Lipschitz network and interval lock enhanced learned index),which con-structs an efficient index through three key mechanisms:(1)BiLipTransformer,a bi-Lipschitz network-based distribution transformer that is inherently injective,incorporates cross-layer connections to capture features even in shallow architec-tures,and allows smooth control of input-output perturbations to improve transformation stability.(2)Training optimiza-tion leverages integral quadratic constraints and Cayley transform-based parameterization to accelerate convergence.Fur-thermore,it adopts the probability integral transform theorem to target a uniform distribution,reducing conflicts in down-stream fitting.(3)PreciseSeekIndex is an index structure combining model nodes and bucket nodes,which uses the interval lock mechanism to resolve global blocking when local insertions trigger retraining.Experimental results demonstrate that,compared with traditional index structure(B+Tree)and representative learned indexes(RMI,ALEX,LIPP,NFL),BLELI achieves an average 2.05×improvement in throughput and a 53.58%reduction in P99 tail latency across multiple datasets and workloads,effectively overcoming performance and stability bottlenecks under complex data distributions and pro-viding a robust solution for efficient index deployment.