100 Million Lines of Code, Millions of $ and thousands of developer hours saved, 0 Wasted Builds; AI Powered Agentic Dev Loop @ Atlassian Scale

Abstract

Managing a massive monorepo with over 100 million lines of code presents significant infrastructure cost, scale, build time and productivity challenges across SDLC.

In the inner dev-loop, this scale results in sluggish search functionality and delayed IntelliSense auto-complete suggestions within the IDE. Additionally, generating repository-level insights and executing deep code refactoring pose significant challenges. To address these issues, an in-house built AI-powered coding assistant enhances productivity in various areas, including feature development, testing, documentation, code review, and maintenance. This tool creates a multiplier effect, alleviating the cognitive load on developers as millions of lines of AI-generated code are seamlessly integrated and accepted by them.

In outer dev-loop, our high volume of daily builds, each running ~500,000 tests, often led to redundant testing. By leveraging static code analysis and machine learning for predictive test selection, we now skip 50-95% of tests per change, dramatically optimizing build times. Further AI-powered tooling has reduced manual effort in code migrations by 50%. These AI/ML advancements in our developer tooling have yielded annual infrastructure cost savings of approximately $4 million and saved thousands of engineering hours.

What’s in it for you?

Discover valuable insights from our extensive experience in effective AI innovation and its applications for managing large repositories. Our goal is to alleviate developer challenges and significantly enhance productivity. This session is tailored for a diverse audience, including software developers, architects, Technical Program Managers (TPMs), and Product Managers, among others.

Some key takeaways:

• Building and scaling AI powered coding assistants even for very large repos
• Predictive Tests Selection using AI/ML
• AI powered diagnostics in inner / outer dev loop and auto fixes
• Fine tuned model for code embeddings