How GitHub renders a 2,200-file, 1M-line pull request without jank

GitHub Blog AI/ML · rss · 2026-09-24

GitHub's engineering blog details rebuilding the pull request view in the Copilot app to handle extreme diffs, stress-tested on an open source PR with 2,200 files, over 1M changed lines, and 400+ inline comments.

Key takeaways:

Pure code diffs rely on an "all heights known before paint" contract: imperative recycled row renderers, typed-array geometry, and backend-streamed diff documents keep per-frame work constant.
Comments break that contract — their height depends on markdown wrapping, expandable blocks, reply composers, and image loading, so estimators fail at the extremes and post-paint corrections cause scroll jumps.
The fix splits geometry in two: exact deterministic code height plus a separate dynamic-block index keyed by identity, with fingerprints and width-bucketed cached measurements.
Their first design (one ResizeObserver per block) was rejected as a self-retriggering feedback loop; the shipped solution is a single idle/scroll-gated measurement pass scoped to 2400px around the viewport, batch-reading mounted blocks in one reflow.

Original post →

How GitHub renders a 2,200-file, 1M-line pull request without jank

More from coding & agent