June 2020 Summaries

The blog post discusses the implementation of an approximate nearest neighbor (ANN) search algorithm in Vespa, focusing on the selection and adaptation of the Hierarchical Navigable Small World Graphs (HNSW) algorithm. It highlights the importance of nearest neighbor search in high-dimensional vector spaces, particularly for real-time applications like image recognition and document retrieval, where the document corpus is dynamic and subject to metadata constraints. The Vespa team evaluated three ANN algorithms—Annoy, HNSW, and RPLSH—based on their performance in indexing and search throughput, with HNSW emerging as the preferred choice due to its superior search efficiency, memory usage, and compatibility with Vespa's real-time update requirements. The post also outlines the challenges of integrating ANN algorithms with Vespa's query tree and filter support, and the necessity of approximate solutions when dealing with large document corpora. The blog concludes with an overview of the decision-making process and the promise of further details on HNSW's integration in subsequent entries.

Big data serving, a complex challenge in applied computing, involves computing over large data sets in real-time, requiring solutions for distributed state management, low latency, high availability, and distributed computation. Traditional methods, like using databases with stateless middle-tier applications, fall short due to network and latency constraints, prompting the need for systems that compute locally where data is stored. Web search, a classic example of big data serving, has driven the development of advanced systems capable of handling the required computational demands. Vespa.ai, an open-source engine developed from the late 1990s web search engine alltheweb.com and later supported by Yahoo!, addresses these challenges by enabling real-time machine-learned model inferences and data organization. With approximately 700 man-years of development, Vespa has become a viable option for various applications beyond search, including recommendation and personalization, offering new possibilities for low latency data computation.