Driving innovation at scale
Powering 41 million daily connections between the digital and physical worlds through a real-time marketplace built for global reliability.
How We Build
Principles forged by the demands of real-time, physical-world systems.
Multi-sided Marketplace
Real-time forecasting, dynamic pricing & matching across earners, consumers and merchants — all in milliseconds.
Hyper-local Geospatial
H3 hex grids, sub-second ETAs, real-time routing and location intelligence powering every trip on the planet.
Platform First
Modular building blocks that let us launch new verticals — rides, eats, freight, autonomous — on shared infrastructure.
Global Adaptability
70+ countries, thousands of regulatory frameworks, dozens of languages — one platform that adapts everywhere.
Resiliency & Scalability
Fault-tolerant systems serving millions of concurrent users. Five 9s is not aspirational, it is mandatory.
Engineered for the Real World
Stories of building the systems that move the world in real time.
Introducing Shard Manager
Even shard counts didn't mean even load — hot Cadence Matching and History hosts overloaded while autoscaled fleets sat idle, and Ringpop's hash ring ignored shard load. Shard Manager, an open-source sharding service, makes ownership explicit with load-aware greedy rebalancing, movement budgets, and etcd-backed state. Hottest-host CPU fell 38%, letting the same traffic run on ~40% fewer hosts.
Read more →Designing a Generalized Ads Marketplace Platform
As Uber Ads grew from Sponsored Listings to sponsored items, display, banner, and video formats, serving logic diverged across surfaces and a growing org struggled with ownership and isolation. A three-layer platform — supply, a product-agnostic marketplace, and demand — normalizes requests via AdSpec. Config-defined fleets and per-provider failure isolation shrink blast radius and speed onboarding of new formats.
Read more →Freight Search: Composing a Shared Search Index from Independently Owned Fragments
Uber Freight search spans 40+ microservices and nearly 100 heterogeneous data sources, so no single team owns the full picture of a load. A declarative fragment framework on OpenSearch lets teams own index slices via Java annotations, with Spark bootstrap and Flink CDC ingestion. 300 fragments merge into 20+ indices, serving thousands of QPS with seconds-level freshness.
Read more →Translations at Scale: Memory Optimization Techniques That Kept Uber's P99 Under 1ms
Uber serves millions of translations per second from memory, but GC ate 20%+ of one service's CPU and cache refreshes doubled memory, causing OOMs. Swapping sync.Map for an RWMutex map, then a sharded hybrid memory/disk store with 6-byte key hashes and LRU eviction, cut CPU 30%+ and memory from 800MB to 40MB with a 99.9%+ hit ratio.
Read more →Ensuring Similar Computational Resources to Stateful Workloads Running on a Diverse Server Fleet
Odin runs ~5M cores across hardware generations, ARM and x86, so identical stateful workloads over-allocate on fast hosts and starve on slow ones. Normalized compute units convert CPU usage through per-host performance factors, then back to physical cores at scheduling time. Rolled out across ~1.4M cores, Kafka allocations dropped ~8% and utilization spread narrowed from 8 to 3 points.
Read more →How Uber Protects Against Retry Storms
Uniform retries turned localized degradation into stack-wide incidents — a failing service three hops deep sees 8× baseline traffic. Uber built error ownership into shared infrastructure: services claim only self-originated errors, and callers retry only claimed ones. During a November 2025 outage it stopped an estimated 9.5M spurious requests, shrinking average retry-storm radius from 20 services to 2.
Read more →Velox at Uber: Rebuilding Presto's Execution Engine for Speed and Scale
Uber's Presto fleet runs hundreds of thousands of daily queries, yet 90%+ of CPU went to scan-filter-project work bottlenecked by JVM overhead. Moving execution to Velox, Meta's vectorized C++ engine, required phased shadow-testing, Parquet fixes, and geospatial extensions. Velox now serves ~90% of eligible traffic — 100PB a day — with ~2.4× faster queries and 3× lower CPU.
Read more →Making Every Core Count: In Pursuit Of Faster Apache Spark Jobs at Uber
Data lake ingestion alone consumes 200,000+ Spark cores, so small efficiency gains matter enormously. Uber offloaded Spark SQL operators to Gluten/Velox with no user code changes, and benchmarked Onehouse's Quanton engine on Hudi write paths. Gluten delivered ~3.5× faster queries at 3.4× lower CPU cost; Quanton hit 5.5× median write speedups with 89% fewer vCPU-hours.
Read more →Running Cost-Efficient Export Workloads at Uber
Narrow export queries like Data Subject Access Requests triggered full scans of huge historical Hudi tables, keeping data hot and blocking GCS auto-class tiering. Rather than secondary indexes, Uber sorted tables on the predicate column and used Hudi metadata column stats to prune files without opening Parquet footers. Sorting cut disk 24.8% and unlocks cheaper Nearline, Coldline, and Archive tiers.
Read more →Scaling Exact COUNT(DISTINCT) for High-Cardinality Non-Rollup Metrics in Distributed Data Pipelines
Exact distinct counts for metrics like quarterly active users meant holding 3.6B identifiers in memory, but monolithic 64-bit bitmaps hit the JVM's 2GB array cap at ~179M uniques. A chunked buffer splits the hash space into 65,536 independent Roaring bitmaps, raising the ceiling to ~11.7 trillion. Two-year backfills dropped 65%, from 8.58 to 2.96 days, with zero OOMs.
Read more →Simplifying Data and Product Integrations with a Data Abstraction Layer
Building a new advertiser report took weeks to months because queries were tightly coupled to constantly evolving data schemas fragmented across sources. Uber's Data Abstraction Layer, an RPC service, maps logical tables to physical ones via declarative config, resolving joins and aggregations across OLAP, Hive, and Docstore. Report time-to-market dropped from multi-week to under two days — roughly 90%.
Read more →Beyond Prediction: Solving the Multiple Knapsack Problem at Scale
Allocating millions of incentive offers across Mobility and Delivery under strict quarterly budgets relied on a manual First-In-First-Out heuristic with no optimization. Tarot frames the problem as a Multiple Knapsack Problem, pairing uplift models and a budget pacer with a CP-SAT solver. Budget utilization climbed from 68% to 99.99%, and 100,000-user problems now solve in minutes instead of 24+ hours.
Read more →How Uber Built an Agentic System to Automate Design Specs in Minutes
Manual design specs across UIKit, SwiftUI, Android XML, Compose, Web React, Go, and SDUI were a bottleneck causing constant documentation drift. uSpec combines AI agents with a Figma Console MCP bridge that reads real tokens and variants directly from Figma — running locally via Cursor over WebSocket. Screen reader specs across all 3 platforms now generate in under 2 minutes.
Read more →Standardized Mobile Analytics for Cross-Platform Insights
Over 40% of Uber's mobile events were ad-hoc custom logs, breaking cross-platform analysis and impression accuracy. The team standardized to three universal event types — tap, impression, scroll — using AnalyticsBuilder classes that capture metadata at the platform layer. Result: 30% less transient-impression noise and reliable iOS/Android parity.
Read more →Unified Checkout: Streamlining Uber's Payment Ecosystem
Every Uber line of business had built payments independently — duplicated logic, inconsistent UX, Apple Pay missing from half the flows. EU Strong Customer Authentication forced a reckoning. Uber built a centralized checkout orchestrator with modular components each LOB plugs into. Holdout results: 3% higher conversion, 4.5% better session recovery, hundreds of millions in incremental gross bookings.
Read more →Measuring Design System Adoption at Scale
Uber's Rider app launches features across hundreds of screens and thousands of feature flags — making manual design system audits impossible. Design System Observability adds a deterministic component scanner that flags non-Base elements, plus a daily screenshot pipeline that auto-files Jira tickets for violations. Teams using Base report 3x faster development and 50% less code.
Read more →Transforming Executive Travel: Delegate Booking
Letting executive assistants book rides for executives meant rethinking trip ownership, identity, billing, and notifications across 30+ backend services and 5 client platforms. Uber introduced a "participant model" extending every booking, tracking, and billing touchpoint to support multiple user profiles per trip — with full audit trails for both EA and executive.
Read more →Uber's Live Activity on iOS
Live Activities run sandboxed with no network access — yet Uber needed real-time driver location on the lock screen. App Groups share on-disk state between the main app and Live Activity, a lightweight DSL syncs content logic across iOS Live Activities and Android push, and an OOA backend debounces updates. Result: 2.26% fewer driver and 2.13% fewer rider cancellations at pickup.
Read more →Scaling AI in Legal: Building Uber's Redlining Agent
Uber's legal teams negotiate thousands of contracts a year, and repetitive redline review slows deals and launches. The Legal Redlining Agent, a Word add-in built on LangGraph, OpenSearch retrieval, and a rules database, recommends accept, reject, or modify actions, drafts comments, and learns from lawyer feedback. It delivers 91% decision accuracy and 20%+ faster contract reviews.
Read more →Designing MCP Gateway: Uber's MCP Management Platform
As teams built MCP integrations independently, tooling fragmented with duplicated infrastructure, poor discovery, and inconsistent security. MCP Gateway centralizes it: AutoCrawler generates servers from Uber's IDL registry, while a proxy translates calls to HTTP, gRPC, or TChannel with tool-level authorization and redaction. It now hosts 800+ servers and 5,000+ tools, with Code Mode as the default.
Read more →Taming the ML Firehose: Scaling Feature Consistency
Training-serving skew — 'en-US' offline versus 'en' online — silently degraded models, and logging 8M QPS of predictions meant ~1.7PB a day. A feature logging framework captures exact inference values as the training source, using allow lists, integer feature IDs, and tuned Flink impression joins. Key-feature mismatch fell from 10%+ to 0%, with payloads shrinking 4–5×.
Read more →Running a Software Factory Efficiently at Uber Scale
Agent usage — and AI spend — exploded as agents took on code review and CI repair, with every turn re-sending growing context. Uber models cost as six multiplied terms and optimizes each: benchmark-driven model selection, compaction, prompt caching, code mode, and a 24M-node context graph. Weekly users grew 7× while cost per session fell 52% from peak.
Read more →Cart Assistant: Agentic Grocery Shopping on Uber Eats
Turning handwritten lists, recipe screenshots, and vague meal plans into a grocery cart forces users to hand-translate intent into dozens of searches. Uber Eats' Cart Assistant decomposes the task into an eight-stage state graph — LLM planning, candidate retrieval, semantic relevance judging, and deterministic price and quantity guardrails — processing items concurrently to keep the agentic shopping flow fast.
Read more →Building a File Semantic Analyzer: Guarding Outbound Data at Scale with AI
Keyword-based data-loss prevention drowned analysts in false positives while missing real exfiltration across thousands of employees' files. Uber's File Semantic Analyzer uses fine-tuned LLMs — OCR, chunking, summarization, and entity extraction — to reason over file meaning and intent, with transparent justifications. False positives fell 97%, incident response dropped from hours to minutes, saving four person-years annually across 150,000 files.
Read more →Securing AI Agents: Lessons From Real-World Deployment
Agents reason over untrusted content and act on it, pushing the attack surface into the semantic layer — prompt injection, tool poisoning, goal hijacking. Uber Secure AI layers provisioning via an AI Registry and Argus code scanning, runtime enforcement with AI Guard and local guardrail models, and continuous red teaming. Explicit intents substantially reduced agents' ability to act on manipulated instructions.
Read more →Solving the Identity Crisis for AI Agents
When AI agents delegate across systems, the original human's identity vanishes, breaking audit trails and fine-grained access control. Uber extended its Zero Trust architecture with an Agent Registry, a Security Token Service issuing short-lived single-hop JWTs carrying the full actor chain, and an MCP Gateway enforcing policy. Deployed across thousands of agents, token exchange stays under 40ms at P99.
Read more →Superuser Gateway: Guardrails for Privileged Command Execution
A misplaced flag in a privileged `rm -r` could silently delete a production dataset with no audit trail. Superuser Gateway removes superuser credentials from engineers' machines entirely, routing privileged commands through a Git-backed PR workflow with CLI submission, CI validation, peer approval, and controlled remote execution. Now standard for all data platform admins.
Read more →Determinism and Safety in IAM Policy Changes
An accidental IAM change on a critical gateway once stopped Uber Eats customers from modifying orders — and ~10% of monthly policy changes involve risky privilege removal. The Policy Simulator pulls 30–90 days of access logs from Apache Pinot and replays them through current vs. proposed policies. Cadence-orchestrated, sub-minute impact analysis before anything ships.
Read more →Automating Kerberos Keytab Rotation at Uber
Rotating 100,000+ Kerberos keytabs is risky: rotation invalidates the previous key immediately, leaving applications without valid credentials. Uber's solution generates keytabs with both old and new versions during transitions, drops fetch intervals to 30s, and integrates with the Secret Management Platform. Now rotates 30,000+ keytabs monthly with zero disruptions.
Read more →Multi-Cloud Secrets Management Platform
150,000 secrets across 25 fragmented vaults — no centralized detection, rotation, or attribution. Uber consolidated into 6 managed vaults, deployed real-time scanning across git/Slack/CI, and built a Cadence-orchestrated Secret Lifecycle Manager. A team of 10 now drives 20,000 automated monthly rotations with 90% fewer secrets exposed in pipelines.
Read more →News
The latest from Uber Engineering and beyond.
Uber Brings ChatGPT Work to All Employees
After a 1,000-person pilot, Uber is rolling out ChatGPT Work with OpenAI to 30,000+ employees — an agent for multi-step knowledge work, wired into Drive, Gmail, Calendar, and Slack.
Read more →Inside Uber's Agentic Pods
Uber embeds engineers alongside operators to hunt for AI wins hiding in plain sight. The results: capital allocation from 15 hours to 30 minutes, financial reports from two days to ten.
Read more →Uber Engineers Named Agentic AI Foundation Ambassadors
Uber engineers Gaurav Goel and Sanjeev Suresh are named Agentic AI Foundation ambassadors — recognizing the technical leadership shaping open standards for how autonomous agents are built, secured, and governed.
Read more →Uber AV Labs at CVPR 2026
Uber AV Labs took its work to CVPR 2026, showing how billions of real trips become training signal for Level 4 autonomy — mining long-tail driving scenarios at a scale few can match.
Read more →Uber at the AI Engineer World's Fair
Fresh from the AI Engineer World's Fair, Uber engineers recap their sessions on the agentic SDLC, GenAI code review, and multimodal agent evaluation — how AI now runs through Uber's engineering.
Read more →Uber GenAI Meetup: Hyderabad
Over 80 AI practitioners, engineers, and builders filled Uber's Hyderabad office for technical talks and a panel on "Beyond the Prototype: Trust, Cost, and Talent in the Age of Agents."
Read more →Research
Browse research and articles authored by our team members.
ADR: An Agentic Detection System for Enterprise Agentic AI Security
Enterprise AI agents built on the Model Context Protocol face limited observability, brittle rule-based defenses, and costly LLM-based detection. ADR pairs agent telemetry, pre-deployment red teaming, and a two-tier online detector — deployed at Uber for 10+ months, screening 10,000+ agent sessions daily.
Continue reading → ICSE-SEIP 2026Scaling Mobile Chaos Testing with AI-Driven Test Execution
Mobile applications in large-scale distributed systems are susceptible to backend service failures, yet traditional chaos engineering approaches cannot scale mobile testing due to the complexity of end-to-end test environments.
Continue reading → SOSP 2025Moirai: Optimizing Placement of Data and Compute in Hybrid Clouds
The deployment of large-scale data analytics between on-premise and cloud sites requires careful partitioning of both data and computation to avoid massive networking costs and performance degradation.
Continue reading →Open Source
Infrastructure for connecting the digital and physical worlds -- open to everyone.
zap
GoBlazing fast, structured, leveled logging in Go.
Go Style Guide
GoThe Uber Go Style Guide — patterns and conventions for writing clean, idiomatic Go.
Base Web
TypeScriptReact Component library implementing the Base design language.
react-vis
JavaScriptData Visualization Components for React.
RIBs
KotlinCross-platform mobile architecture framework powering the Uber rider and driver apps.
fx
GoDependency injection based application framework for Go.
Kraken
GoP2P Docker registry capable of distributing TBs of data in seconds.
H3
CHexagonal hierarchical geospatial indexing system.
CausalML
PythonUplift modeling and causal inference with machine learning algorithms.
goleak
GoGoroutine leak detector for Go.
automaxprocs
GoAutomatically set GOMAXPROCS to match Linux container CPU quota.
ratelimit
GoBlocking leaky-bucket rate limit implementation.
dig
GoReflection based dependency injection toolkit for Go.
NullAway
JavaEliminate NullPointerExceptions with low build-time overhead.
ADR
PythonSecures enterprise AI agents through observability, security benchmarking, and threat detection. Deployed at Uber.
Michelangelo
PythonUber's end-to-end machine learning platform for building, training, and serving models at scale.
Ready to build?
Join engineers across 4 continents and help bridge code and the physical world.
Where would your code empower real humans?
View Open RolesSenior ML Engineer – AV Foundation
Drive frontier ML research for autonomous vehicles — design novel modeling approaches and own the full lifecycle from data preparation through deployment. Based in Sunnyvale.
Apply now → Applied AIStaff ML Engineer – Applied AI
Define and lead foundation model strategy for AI-powered discovery across Mobility and Delivery — shape technical direction and mentor senior engineers at global scale.
Apply now → SecuritySenior Security Incident Commander
Lead Uber's critical security incidents end-to-end — balance real-time command decisions with deep technical investigation, then drive systemic fixes through post-incident analysis.
Apply now →