Catalogue / uber-2026-eats-search-pipeline

Uber Eats halves search latency across presentation, retrieval, hydration, ads serving, and Go/mesh infrastructure

Uber, 2026. Uber Eats search pipeline.

Organisation
Uber
Evidence
Reported from production
Source
Halving the Time: How Uber Eats Rebuilt Its Search Pipeline

Uber Engineering Blog, published 2026-09-11

Records

Decision#d1

Measure Above-the-Fold completion, paginate with a server-side cache, and render HTML templates concurrently

Decision#d2

Drop low-yield lexical retrieval arms, deduplicate chains earlier, and group features by product embeddings

Decision#d3

Split monolithic hydration into parallel ranking and presentation phases and remove false DAG dependencies

Decision#d4

Hedge slow requests across four presentation hydration dependency layers to reduce tail latency

Decision#d5

Move ads bid scoring to a columnar layout with in-memory ad data and drop redundant serialization

Decision#d6

Fix encoding, embedding size, service-mesh connection caps, and Go GC overhead found via profiling

Sign in to read what was chosen, what was rejected and what it cost, each with its place in the source. An account is free.