Cadence: Writing Workflow Code as If Failures Never Happen
How Uber's Cadence engine lets developers write long-running workflow logic as plain code while the platform handles retries and state durably.
Engineering at scale
14 articles · 55 min of reading
How Uber's Cadence engine lets developers write long-running workflow logic as plain code while the platform handles retries and state durably.
How Uber's DeepETA replaced purely graph-based routing calculations with a deep learning model to correct systematic ETA errors at low latency.
How Uber built a payments platform to handle rider charges, driver payouts, taxes, and refunds correctly across dozens of countries and currencies.
How Uber built Michelangelo to give data scientists a shared path from feature data to a deployed, monitored production ML model.
How Uber computes and updates dynamic surge pricing in near real time by combining geospatial indexing, streaming demand signals, and ML forecasting.
How Uber ended up running a polyglot backend split mainly between Go and Java, and why it consolidated each language's services into monorepos.
How Uber built M3, its open-sourced metrics platform, when off-the-shelf time-series tooling couldn't handle its monitoring scale.
Why Uber built Schemaless, an append-only document layer on top of sharded MySQL, instead of migrating wholesale to a NoSQL database.
How Uber built RAMEN, a persistent-connection push platform, to replace polling and cut latency and backend load across its rider and driver apps.
How Uber built Hudi to bring record-level updates, deletes, and incremental processing to a Hadoop data lake designed for immutable batch files.
How Uber's Ringpop library used consistent hashing and a gossip protocol to shard stateful services without depending on a single coordinator.
Why Uber built a hexagonal hierarchical grid system to power surge pricing, ETAs, and driver-rider matching, and open sourced it as H3.
How Uber's Domain-Oriented Microservice Architecture organized thousands of services into gateways and layers after sprawl made the old model unmanageable.
The write amplification, replication, and connection-handling problems that pushed Uber off Postgres and onto a MySQL-based storage stack.