Harness Rebuilds Git for AI Agent Scale: Architecting for High-Throughput Code Generation
Harness rebuilt its Git repository to handle the immense load generated by AI coding agents, which produce significantly more pull requests and commits than human developers. This article explores the architectural shift from a human-centric Git model to an AI-first, scalable repository designed for continuous, high-volume code changes, emphasizing the challenges and solutions in continuous integration/continuous delivery (CI/CD) pipelines.
Read original on The New StackThe Challenge: AI Agent Traffic Overwhelming Traditional Git
The advent of AI coding agents like GitHub Copilot and Amazon CodeWhisperer has drastically increased the volume of code changes, leading to a 1.5x to 50x surge in pull requests (PRs). Traditional Git platforms, designed primarily for human teams (10-15 engineers) and PRs lasting hours to days, are buckling under this sustained, high-frequency, 24/7 agent traffic. This overload creates significant bottlenecks in testing and review processes, forcing teams to either lower risk tolerance or face unmanageable backlogs.
Harness's Architectural Response: An AI-First Git Repository
To address this, Harness rebuilt its open-source Git repository as an "AI-first" platform, specifically engineered for both human and AI traffic. Key architectural considerations for this new system include:
- Kubernetes-based Architecture: The new repository leverages Kubernetes for orchestration, enabling a distributed and scalable infrastructure.
- Multi-Cloud and Multi-Region Deployment: Designed to run across various cloud providers and geographical regions, enhancing resilience and global availability.
- High Throughput Testing: Tested to handle thousands of commits per second, a significant leap from the average ~1,000 commits per second GitHub handles across its entire platform.
- Software Delivery Knowledge Graph: Harness also developed a knowledge graph mapping pipelines, deployments, incidents, and policies. This provides AI-driven context for code reviews, helping prioritize important changes (e.g., actual logic changes) over boilerplate or dependency updates, saving significant manual review time.
Scaling Considerations for Version Control
When designing a version control system for high-throughput environments, consider the following: distributed data storage, eventually consistent replication strategies, efficient indexing for search and metadata, and a robust queuing system for processing continuous commit streams. Traditional relational databases might struggle; NoSQL or custom-built solutions optimized for write-heavy workloads may be necessary.
Deterministic Tooling and the Future of CI/CD
Despite the rise of autonomous agents, the article emphasizes that certain parts of the CI/CD pipeline should remain deterministic. This includes relying on existing test runners for validated test results, rather than replacing them entirely. The focus shifts to enhancing existing deterministic tooling with AI capabilities, particularly for optimizing the human review process by highlighting critical changes and filtering out noise (like dependency updates). The ultimate goal is an autonomous delivery lifecycle, but organizations must be ready for such a paradigm shift, starting with incremental improvements in review efficiency.