Stripe’s engineering blog: from service mesh replacement to 25M-line refactors
Stripe’s engineering blog has chronicled the company’s work across infrastructure, developer productivity, AI, and payments. The latest posts highlight how the company handles scale challenges that few other organizations face — from processing traffic above 1.6% of global GDP to maintaining a 50-million-line Ruby monorepo.
Infrastructure at Stripe’s scale
One of the most substantial recent posts details the replacement of Envoy with a custom-built distributed proxy. With a reliability target of 99.9995% and traffic that has grown past 1.6% of global GDP, the infrastructure team found the popular service mesh insufficient. The post walks through the architecture of the new data plane built from scratch.
Database operations have also been automated extensively. Stripe modeled its MongoDB infrastructure as a traversable graph, then used pathfinding algorithms to compute and execute recovery plans dynamically. The results: pager volume dropped by 30% (roughly 200 pages per year), and unhealthy shard states that previously caused 12 days of downtime annually were eliminated. The system also supports new shard layouts with no manual intervention.
Another write-up covers Stripe’s document databases, which sustained 99.999% uptime while performing zero-downtime data migrations. The Data Movement Platform is the centerpiece, and the post explains how it was designed and applied across the fleet.
For billing, Stripe introduced real-time analytics to help businesses spot customer trends as they emerge. The feature was built in response to a survey finding that 84% of global business leaders believe rapid pricing adaptation will be a key competitive advantage. Separate posts cover the jurisdiction resolution system (JRS) for Stripe Tax — a faster, less resource-intensive way to determine tax obligations in regions with overlapping jurisdictions — and the Ledger system for tracking and validating money movement.
Stripe has also begun extending its Global Payments and Treasury Network (GPTN) with stablecoins. The network is described as a graph with accounts as nodes and payment rails as edges; it batches transfers to reduce cost, nets opposing flows, and routes through optimal paths. The post on this topic examines why most money movement still happens in batches — with wires settling at end of day and ACH taking one to three business days — and how the network handles these constraints.
Developer productivity at massive codebase scale
Several posts focus on tooling for Stripe’s enormous codebases. The Developer Productivity team extended rubyfmt, a Rust-based autoformatter, and rolled it out across a 25-million-line Ruby codebase in a single overnight operation. The tool is zero-config and ultra-fast.
For CI, Stripe’s Selective Test Execution system allows the company to keep scaling its team and codebase while running only about 5% of its tests on average. The post explains the clever tricks used to achieve this in a 50M-line Ruby monorepo.
Java modernization is another focus. Stripe outlined how it built an upgrade path for large JVM codebases, addressing the common pain point where recurring upgrade costs leave most organizations trailing years behind the latest Java versions.
Stripe also shared how it converted more than 3.7 million lines of code to TypeScript in a single pull request. On a Sunday in March 2022, the migration went through, and hundreds of engineers began writing TypeScript the next day. Another post covers Sorbet, Stripe’s type checker for Ruby, which is used across 15 million lines of code. A separate piece details how API changes flow from engineers’ pull requests through Stripe’s large developer product suite to keep everything in sync.
On the security side, one post explores the tradeoff between fast builds and secure builds. Stripe’s CI system orchestrates pipelines and runs tens of thousands of test suites, relying on a mix of open-source technology and custom engineering to meet a high security bar. Another post covers fast but secure builds in depth — the title says it best: “Fast builds, secure builds. Choose two.”
AI agents and internal platforms
Minions are Stripe’s homegrown coding agents, responsible for more than a thousand pull requests merged each week. Humans review the code, but the agents write it end to end. Two posts (part 1 and part 2) explain how they work and how they were built.
The Knowledge AI Platform is broader in scope — an AI agent platform for non-coding knowledge work. It handles tasks from quick queries to multi-day projects, connecting employees to over 1,000 internal tools and skills for enterprise-scale productivity.
Stripe also published details on Shepherd, its adaptation of the open-source Chronon system to scale ML feature development. The post outlines the technical architecture and how Chronon’s capabilities were expanded to meet Stripe’s scale.
Fraud, payments tooling, and docs
Stripe Radar, the company’s fraud prevention solution, is the subject of a post covering the key decisions and lessons learned over its years of development. A separate post on Stripe Credits describes how the team built a programmable, auditable virtual payment method for prepaid and Stripe-issued credits that integrates with accounting and compliance frameworks.
Test clocks made it easier to test Stripe Billing integrations, and a post explains how they were built. For documentation, Markdoc is Stripe’s open-source format that delivers a good user experience without compromising authoring — it lets writers express interactivity and page logic without mixing code and content.
Looking back at API design, Stripe’s payments APIs have evolved over a decade with a focus on abstracting away payments complexity. The retrospective post covers the first 10 years of that evolution.
