March 2021 Summaries

Vespa, an open-source big data processing and serving engine largely developed by Yahoo engineers, has released several updates in March 2021, enhancing its functionality and efficiency. These updates include new features in the /document/v1/ API, which now allows users to update or remove multiple documents and copy them between clusters, making the API more expressive and useful for both production and learning purposes. The introduction of weakAnd.replace enables automatic conversion from OR to WeakAnd in queries to reduce latency by focusing on the most relevant hits. Additionally, improvements in feed-block management and staggered memory flushing during stop/restart processes aim to prevent index corruption and reduce temporary memory usage, respectively. These enhancements are part of Vespa's ongoing development, supported by community feedback and contributions, and are used in various Yahoo and Verizon Media products.

The blog post by Thiago Martins on the Vespa Blog details the process of transforming a news search application into a news recommendation app using embeddings with the Vespa engine. It explains how embeddings are used to represent both users and news articles as vectors, enabling recommendations through an approximate nearest neighbor (ANN) search. The guide covers the setup of user and news schemas with dense tensor fields for embeddings and describes how Vespa's ANN search can be combined with query filters to refine results. The post also includes steps for indexing news embeddings with the HNSW algorithm, deploying the updated application, and using Python functions to query user embeddings and retrieve personalized news recommendations. Additionally, it highlights Vespa's capabilities in integrating ANN search within its query tree, allowing seamless filtering without extra processing steps.

Thiago Martins, a data scientist at Vespa, outlines the process of building a news recommendation app using Vespa within a Python environment in the first part of a tutorial series. This initial installment focuses on creating basic news search functionality, leveraging the demo version of the Microsoft News Dataset (MIND) for demonstration purposes. The tutorial provides a step-by-step guide on setting up the app, including installing the necessary tools like pyvespa, creating an application package, adding fields to the schema, and deploying the app using Docker. The article also explains how to feed data into the app and query it using Vespa's query API, enabling users to perform searches and sort results based on relevance scores. Additionally, it introduces an enhanced ranking feature that incorporates article popularity, which is determined by the ratio of clicks to impressions, although it acknowledges that this simplistic measure could be improved with machine learning in future implementations.

The blog post by Kristian Aune from Vespa.ai explains how Postman can be used to streamline the process of working with mTLS-secured Vespa APIs, which can otherwise be cumbersome due to the complexity of handling JSON APIs. It highlights the benefits of using Postman to manage credentials, endpoint URLs, and modify JSON requests, which facilitates running queries, as well as POST/PUT/DELETE operations on Vespa instances. By importing Postman collections and configuring certificates for secured endpoints, users can efficiently test and manipulate data without the hassle of manual command-line operations. The post emphasizes the ease of sharing Postman collections across teams, enhancing collaborative efforts and allowing developers to focus more on data and schema management rather than on the technicalities of API interactions.

The blog post introduces pyvespa, a Python API for Vespa, highlighting its utility in facilitating fast prototyping and machine learning experiments for Vespa applications. It outlines three primary methods to leverage pyvespa: connecting to a running Vespa application, building and deploying an application using the pyvespa API, and deploying from Vespa configuration files stored on disk. The article provides examples of each method, such as connecting directly to an existing Vespa application or building an application from scratch using the pyvespa API, which can then be deployed to a Docker container. Additionally, it discusses the possibility of deploying applications from existing Vespa configuration files, emphasizing the library's role as an experimental tool for information retrieval rather than a production-ready solution. The flexibility offered by these methods allows for easy experimentation while maintaining the capability to incorporate features not directly available through pyvespa.