Performance

The short version: steady frames cost about the same for 1,000 lines as for 100,000, because rendering only paints visible rows. Comparison, first layout, and resize scale with the document, so large inputs belong on a worker or behind a loading state.

Guidance for hosts

• Prepare off the event loop. DiffDocument::from_text, compare, and parse are synchronous. So is syntax preparation.
• Expect the first render and each resize to touch the whole document. They build glyph and row indexes. Later frames reuse them.
• Cheap per-frame changes: theme, word highlights, sticky headers, scrolling, and attaching new syntax styles reuse layout.
• Changes that relayout: width, wrapping, mode, and folding.
• Search updates scan the retained source once per query change; rendering reuses the matches.

Measurements

Apple M2 Max, Rust 1.99, release build, 100×30-cell buffer, October 9, 2026. Values are the range of two short Criterion runs’ mean estimates, split view with wrapping. They show where costs are, not latency guarantees.

Operation	1,000 lines	100,000 lines
Compare (insertion fixture)	0.15 ms	16 ms
First layout	1.2 ms	135–159 ms
Resize	0.14 ms	17–19 ms
Toggle one file fold	0.12–0.19 ms	15 ms
Search update, both sides	0.06–0.08 ms	10 ms
Steady frame	75–80 µs	89–177 µs
Hit-test one cell	15 ns	15 ns

Steady frames were 68–82 µs across unified and split, wrapped and unwrapped, at both sizes. Closed files redraw in about 23 µs.

Memory is dominated by prepared layout: the 100,000-line Unicode fixture uses about 33 MiB after comparison and about 650 MiB resident after its first layout. Work to shrink the glyph representation is tracked in the repository’s performance notes.

Reproduce

cargo bench --bench viewer --locked -- \
  --warm-up-time 0.1 --measurement-time 0.2 --sample-size 10
RATATUI_DIFF_MEMORY=1 cargo bench --bench viewer --locked

The performance notes list what each benchmark includes and excludes, the stress fixtures, and the full result tables.