Sharding Strategies for Scalable Vector Databases

Within the scope of this study, a systematization and comparative analysis of various sharding strategies applicable to distributed vector stores have been carried out.The objective of the work is to analyze strategies for the joint use of scalable vector databases.The methodological approach is based on a review and synthesis of advances in the architectures of distributed vector databases and their partitioning techniques as reflected in scientific publications and technical reports.To assess the effectiveness of each strategy, a comparative analysis was conducted according to criteria such as access latency (latency), throughput (throughput), search accuracy (recall/precision), and load balancing (load balancing).The scientific novelty of the study lies in the proposal of a multi-criteria framework decision model that takes into account the heterogeneity and variability of workloads in modern AI systems.Practical conclusions demonstrate that hybrid strategies providing an adaptive combination of data partitioning with query processing parallelism exhibit the greatest versatility and efficiency when operating with highly loaded vector databases.The results obtained are of significant interest to data engineering practitioners, scalable system architects, and researchers in the fields of data management and artificial intelligence.