Vespa’s position in the ML ecosystem
Posted by Lester Solbakken on 4. februar 2019
In recent years there has been significant developments and breakthroughs in analyzing and learning from large amounts of data. Particularly the advances in deep learning has arguably been the primary driver for this, as the bar has been raised in many fields such as image recognition and textual analysis. For the general public, perhaps the most evident has been the highly publicized success of AlphaGo, AlphaZero and now AlphaStar, which has demonstrated the advances in AI and reinforcement learning. At Verizon Media, the insights gained from analyzing data is commonly used to improve our websites in some manner. For instance, machine learned models are used to successfully increase personalized relevance in search results or ad impressions. In other cases, such as in recommendation systems, machine learning is the core technology.
With the advances on the algorithmic side of machine learning there has naturally enough been a corresponding increase in learning frameworks as well, such as TensorFlow, PyTorch/Caffe2, and MxNet. While many of these frameworks are relatively easy to set up and use for training models, model inference in production is leess straight forward as it depends heavily upon the concerns of the application as a whole. In larger applications, models are not usually run in isolation, but as a part of a system collaborating to compute relevant pieces of information. This presents some unique and hard challenges when it comes to engineering solutions that work at scale
In this talk we present Vespa, which is an open source platform developed at Yahoo for building applications that carry out scalable real-time data processing over large data sets. Vespa has rich capabilities for performing general computation, including features for machine learned model inference. We will talk about Vespa’s position in the ML ecosystem, how performance at scale is achieved, and explore the challenges in building web-scale data and ML driven applications. We present some real-world examples running on Vespa, including an application using reinforcement learning to rank comment fields.
For more information, check vespa.ai. If you want to get your hands wet, check out the quick start at https://docs.vespa.ai/documentation/vespa-quick-start.html or delve deeply into the tutorial at https://docs.vespa.ai/documentation/tutorials/blog-search.html.
Lester Solbakken is a principal software engineer at Verizon Media (formerly Yahoo) where his focus is on machine learning solutions on Vespa.
Lester Solbakken
Principal Software Engineer at Verizon Media (formerly Yahoo)