2026/04/30 by Geoffrey X. Yu, Ryan Marcus, Tim Kraska
Computer Science · #cs.DB
paper · pdf · doi:10.48550/arxiv.2604.28079
18 pages, 15 figures
arxiv created 2026/08/06 · arxiv updated 2026/08/07
Relational database management systems (RDBMSes) can process general-purpose queries, but often have lower performance compared to custom-built solutions for specific queries. For example, consider a group-by query over a few known groups (e.g., grouping by country). While an RDBMS would likely use a hash map to do the grouping, a faster method could hard-code the expected groups into the query executor. Such workload-specific techniques, which we call query accelerators, are not widely used in practice because the engineering effort (optimizer and engine changes, potential bugs) does not always justify the isolated performance gains (speedup on a specific query). We propose Tailwind: a non-invasive query planner that brings accelerators into any RDBMS that supports data import/export. Accelerator builders register accelerators using abstract logical plans (ALPs): a new abstraction based on regular tree expressions that specifies the logical sub-plans each accelerator can correctly replace. Tailwind also uses each ALP's structure to automatically build a neural network model to predict the accelerator's performance. At runtime, Tailwind sits atop an RDBMS and transparently rewrites queries to run across one or more accelerators when predicted to be beneficial, falling back to the underlying RDBMS when not. Across three distinct case studies, we use Tailwind to integrate workload-specific accelerators with Redshift and DuckDB to achieve geomean speedups of 1.38x, 1.76x, and 1.28x.