Daniel Krefl: Hacking pgvector for performance скачать в хорошем качестве

Daniel Krefl: Hacking pgvector for performance

In this talk I will present a modification to the popular pgvector Postgres extension for approximate nearest neighbor search. The modified extension keeps the vector index persistent in external (Non-Postgres) memory, and pushes potential filter conditions directly into the index scan. Storage of index and compute can also be offloaded to a GPU. I will explain in detail the inner workings of a Postgres index scan, the rational behind moving core parts of pgvector away from Postgres, and the performance implications. I will also highlight how this modified extension finds application in massive data processing within the framework of the AERO project for the future heterogeneous EU cloud infrastructure.