April 2026 Summaries

The blog post by Kai Borgen, a Technical Product Engineer, explores scaling a Vespa application to handle enterprise-scale workloads using the MS_marco passages dataset, a comprehensive open dataset for information retrieval. Vespa is described as an AI search platform and an all-in-one solution for retrieval and large-scale computation needs. The post details the process of feeding the extensive dataset into a Vespa application, starting with minimal resources and progressively scaling up to a configuration with 100 GPU container nodes and 40 content nodes. This scaling process highlights the importance of identifying and addressing bottlenecks, particularly in GPU and CPU utilization, to optimize performance. The ultimate goal is to achieve high feed throughput, reaching approximately 7358 documents per second, showcasing Vespa's capability to process over 8.8 million passages in just over 20 minutes. The blog emphasizes the use of Vespa Cloud's metrics dashboards to monitor performance and make informed decisions on resource allocation, underlining the adaptability of Vespa to different computational requirements.

The Vespa Cloud Metrics Dashboard is designed to streamline the process of troubleshooting production issues by guiding users through a systematic approach of identifying system health, latency sources, and resource constraints. It simplifies monitoring by focusing on key questions about system health, bottleneck identification, and resource utilization, allowing users to move efficiently from symptoms to actionable insights. Recent updates to the dashboard include the addition of Health Indicators, annotations for service restarts and core dumps, and a detailed breakdown of container thread pools and JVM memory, enhancing its capability to provide context and clarity during operational events. By using a structured workflow, the dashboard helps users quickly pinpoint bottlenecks and connect latency spikes to specific operational changes, making it a practical tool for diagnosing and resolving performance issues.