DEVOURED
Halving the Time: How Uber Eats Rebuilt Its Search Pipeline
Uber cut Eats search latency by 50% through massive architectural simplification and by utilizing AI coding agents to automate performance optimizations.
What: Uber engineers reduced search pipeline latency by parallelizing hydration stages, removing redundant lexical retrieval, adopting column-oriented storage for ads, and using Go value types to reduce garbage collection overhead. They also used an agentic workflow to identify and apply these optimizations.
Why it matters: This demonstrates a shift in engineering culture where 'sweating the details'—such as memory allocation and serialization—is now driven by automated AI analysis of production profiles.
Takeaway: If you are facing high GC overhead in Go services, switch your data models to value types instead of pointers to shift allocations from the heap to the stack.
Deep dive
- Reduced retrieval latency by removing low-value lexical paths and leveraging product-level embeddings.
- Optimized ranking by splitting hydration into parallel ranking and presentation phases.
- Implemented column-oriented data layouts for ad bids to minimize serialization overhead.
- Improved Go performance by switching pointer-heavy data models to stack-allocated value types.
- Used request hedging to mitigate tail latency in hydration dependencies.
- Utilized an agentic pipeline to identify bottlenecks in live profiles, draft PRs, and validate fixes with benchmarks.
Decoder
- Hydration: The process of populating a search result with metadata (pricing, images, stock status) after the initial ID-based retrieval.
- Request hedging: A strategy of sending identical requests to multiple backend instances and using the fastest successful response to reduce tail latency.
- Lexical retrieval: Searching based on text matches (e.g., keyword searches) as opposed to semantic or vector-based search.
Original article
Full article content is not available for inline reading.
Explore all 67 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi