Uber Eats Reduces Search Latency by 50% Through Pipeline Rebuild

Uber Eats Achieves Significant Latency Reduction

Uber has announced a 50% reduction in end-to-end search latency for Uber Eats. This improvement stems from a comprehensive rebuild of major parts of the search pipeline, affecting components such as retrieval, feature hydration, ranking, advertising, and underlying infrastructure.

Shift in Performance Measurement

A key change in this optimization effort was the shift in the primary latency metric. Instead of focusing on backend API response time, Uber began measuring 'Above-the-Fold completion,' which tracks the time until the initial screen of search results, including images, is fully rendered. This new metric provided a more user-centric view of performance.

Key Technical Optimizations

Several technical changes contributed to the latency reduction. Pagination with server-side caching and asynchronous rendering improved Above-the-Fold latency by over 200 milliseconds. Retrieval work was reduced by removing low-value strategies and using product-level embeddings, saving approximately 170 milliseconds. Separating ranking hydration from presentation data and redesigning the advertising path with column-oriented bid data further contributed to the overall latency reduction.

Engineering Principles and Approach

Engineers involved in the project emphasized that the improvements resulted from a series of incremental optimizations rather than a single large architectural change. The approach focused on 'doing less work' and 'avoiding unnecessary waiting,' following a 'Measure, Identify, Fix, Validate' loop for continuous performance enhancement across the full stack.