Redesigning source control at Anthropic (anthropic.com)

Article summary

Anthropic redesigned their source control system to support a large increase in code changes, from 2,500 to 30,000 merges per day, by building an in-house version control system called Scion. Scion is designed to work with always-online agents and jobs, and it uses a client-server architecture with a server that answers questions about file history and blame. The new system has improved performance and scalability, reducing the infrastructure footprint by 200x and speeding up code checkouts by two orders of magnitude. This change was necessary due to the rapid growth of their monorepo, which was straining their previous git-based system.

Main themes

• source control
• scalability
• AI-generated code
• monorepo management
• version control systems
• software development lifecycle

What commenters say

• The use of AI-generated code may lead to a shift in how source control systems are designed and used, with a focus on storing prompts rather than code.
• The idea of storing prompts as source code is not practical due to the complexity of software development and the need for nuance in specifications.
• The development of new version control systems like Scion may be a response to the limitations of traditional systems like git in handling large, rapidly changing monorepos.
• The use of AI models to optimize code and improve performance can be beneficial, but it also raises concerns about skill atrophy and overreliance on these models.
• The design of Scion, with its client-server architecture and focus on always-online agents, may be a model for future version control systems that need to support large-scale, distributed development.
• The question of whether it is better to store code or specifications as the primary source of truth in software development is still an open one, with different approaches having different advantages and disadvantages.