June 2018 Summaries

Vespa, an open-source big data serving engine, has introduced support for the Open Neural Network eXchange (ONNX) format, allowing for seamless evaluation of machine learning models from various frameworks with low latency over large datasets. This integration complements Vespa's existing TensorFlow support and enables interoperability among popular deep learning frameworks like TensorFlow, PyTorch, and MxNet by using ONNX as a common intermediate representation in model deployment and development. Vespa optimizes model evaluation for real-time responses across large datasets and aims to streamline the process of deploying models trained on different frameworks. Importing ONNX models in Vespa involves adding the model to the application package and referencing it using a new ONNX ranking feature. While Vespa currently supports many ONNX operations, some complex operations like convolutional neural networks and recurrent networks are not yet fully supported due to computational costs. Vespa continues to enhance its performance and expand ONNX support, including the ONNX-ML extension, and invites feedback from users to further improve its capabilities.

Vespa's new parent-child feature allows for efficient modeling of hierarchical data relationships, providing a middle ground between the traditional relational database approach and the scalability-focused NoSQL model. This feature is particularly beneficial for cases where data can be represented as a directed acyclic graph, such as in a hypothetical startup serving cat picture advertisements, where Advertisers have multiple Campaigns, each with multiple Ads. Parent-child relationships enable faster updates across related data points, like campaign budgets, by using reference fields and imported fields instead of conventional joins, which are challenging to scale. While this approach incurs a minor performance penalty during search queries, it significantly speeds up data updates and avoids the complexities of maintaining denormalized data models. Vespa's implementation contrasts with ElasticSearch by offering real-time parent-child reference management without the need for manual shard management or experiencing significant query slowdowns. Despite some limitations, such as the inability to query parents based on child properties, Vespa's parent-child feature is a powerful tool for applications requiring scalable, hierarchical data modeling.