TechPrufer
Get started
GitHub
GitHub logoGitHub · Feature trail

Pull Request Files Changed Performance

4 posts · Dec 2016 – Sep 2026

What this trail covers

This feature thread tracks the evolution of the Pull Request 'Files changed' tab performance. Initially, the focus was on general responsiveness. This post details significant optimizations for rendering diff lines, reducing DOM node counts, memory usage, and improving Interaction to Next Paint (INP) scores, especially for large pull requests. It introduced strategies like focused diff-line optimizations, graceful degradation with virtualization, and foundational rendering improvements. The v2 release of the 'Files changed' tab introduced a new architecture to handle extremely large pull requests by splitting the document's height into deterministic code geometry and dynamic block geometry, with a sophisticated measurement scheduler and scroll anchoring to maintain responsiveness.

All GitHub trails
  1. Rendering huge pull requests in the GitHub Copilot appThis post details significant optimizations for rendering extremely large pull requests in the GitHub Copilot app's pull request view. It addresses the challenges of rendering large diffs with numerous comments by introducing a dual-geometry approach: deterministic code geometry and dynamic block geometry. Key contributions include a scroll-gated measurement scheduler that prioritizes on-screen reads, off-screen measurement fallbacks, and the use of ResizeObservers that flag blocks for re-measurement rather than directly triggering layout updates. It also introduces scroll anchoring to correct viewport jumps without fighting user interaction and distinguishes between user-initiated scrolls and programmatic scrolls to prevent visual glitches. The core problem solved is maintaining a performant review experience for pull requests with thousands of files and millions of changed lines.
  2. The uphill climb of making diff lines performantThis post details the engineering effort to optimize the 'Files changed' tab performance in pull requests, particularly for large ones. Key contributions include: reducing the number of React components per diff line from 8 to 2, simplifying the component tree, moving complex state (comments, context menus) to conditionally rendered child components, adopting O(1) data access using JavaScript Maps for state lookups, restricting useEffect usage to top-level components, and implementing window virtualization for rendering only visible portions of massive pull requests. These changes resulted in significant reductions in total lines of code, unique component types, components rendered, DOM nodes, memory usage, and INP scores.
  3. Crafting a better, faster code viewThis post details the development of a new code view architecture for GitHub. It addresses performance issues with large files by introducing an invisible textarea containing the entire file content (for accessibility and native find functionality) and a virtualized, syntax-highlighted overlay for visual display. This approach reduces DOM nodes, improves LCP and TTI, and enables native browser find capabilities, overcoming limitations of previous virtualization attempts.
  4. How we made diff pages three times fasterThis post details a significant refactor of how diff pages are rendered, moving from a monolithic approach to a progressive loading strategy. It introduces the use of `git-diff-tree --raw` to generate a 'table of contents' for diffs, followed by `git-diff-pairs` to fetch patch text for individual files. This approach significantly reduces initial load times, improves accuracy of diff statistics by using `git-diff-tree --numstat --shortstat`, and allows for more flexible handling of large diffs by setting new, higher effective limits for the entire diff while loading smaller batches initially. The post also details the use of a Scientist experiment to compare the old and new approaches.