0
Fork 0
mirror of https://github.com/zeronsh/comet.git synced 2026-11-16 18:57:19 +00:00
comet/docs/performance-runway-layout.md
Wing b004883c44
Fix runway scrolling regressions and release v0.2.42 (#261)
* Fix runway layout, smooth sends, and streaming scroll handoff

* Run UI layout regression tests in pull request CI

* Record matched runway benchmarks and prepare v0.2.42

---------

Co-authored-by: wing-anara <wing@anara.com>
2026-09-05 23:30:15 +00:00

4.9 KiB

Runway layout regression and performance validation — PR #261

The layout fix showed no material CPU or memory regression in eight matched Linux replays against v0.2.41 (489a8c2). The tested candidate is d8c5760, including that same sidebar fix. Streaming CPU changed by −0.26% on long replies and +0.48% on short replies. Peak RSS differed by less than 0.8 MiB in both cases. The short-stream UI main-thread difference was +0.29 percentage points; this comparison does not claim literally zero additional work or a universal speedup.

The GPUI dependency changes from 36ca993 to 8425850, whose only change is virtual-list reservation sizing. Renderer, scene caching, shared snapshots, markdown caches, and spring parking optimizations are preserved. Reservation layout visits the existing measured window and queries the height tree; it does not render the full transcript or add an idle animation loop.

Workload Metric v0.2.41 Candidate
Long Streaming CPU (%) 136.19 135.83
Long UI main-thread CPU (%) 10.25 10.25
Long Peak RSS (MiB) 220.73 221.47
Long Peak PSS (MiB) 163.33 164.08
Long Settled CPU, last 10 s (%) 7.78 7.78
Short Streaming CPU (%) 95.83 96.30
Short UI main-thread CPU (%) 7.28 7.57
Short Peak RSS (MiB) 209.53 210.32
Short Peak PSS (MiB) 152.75 153.29
Short Settled CPU, last 10 s (%) 5.73 5.68

Values average two runs per build, in main/candidate/candidate/main order for each workload. CPU is percent of one core. Memory sums UI and engine phase peaks, which may occur at different instants. All runs completed with identical reply hashes within each workload. The final ten settled seconds show no sustained idle CPU increase. Individual summaries and executable/source identities are in the raw results.

Functional regression coverage

All 611 release-mode UI tests passed, including ten new headless GPUI layout and handler tests. The initial new tests failed on v0.2.40/main: the short-chat animation jumped from 660 px to 0 px in one frame, and a three-row append increased the scroll range from 2 px to 230 px before the corrective frame.

The new coverage checks:

  • Smooth, monotonic short-chat glides and first echoes arriving after send.
  • First-paint geometry when multiple rows consume a runway.
  • Real transcript streaming through overflow, tail-following, and completion.
  • Real wheel events at the end, without entering a gap or reversing direction.
  • Background commits without frames, followed by downward-only navigation.
  • Second sends and steering while preserving the viewport until echo arrival.
  • User scroll-up ownership while new output outgrows the reservation.
  • Viewport resizing in the same layout, without a provisional gap.
  • Direct navigation to a virtualized tail with unknown preceding row heights.

Existing regressions also cover same-layout completion shrinkage, remeasurement during glides, off-screen overflow retirement, session restoration, and stale hold cancellation. The UI suite now runs in .github/workflows/ui-tests.yml for relevant pull requests and main updates.

Native recordings and measurement scope

The PR has two GitHub user-attachment videos from a native Linux/X11 window: first send → runway → overflow following → completion → repeated downward input; and second send → scroll up → focus another window during streaming → refocus and use only downward input. They were captured before integrating the unrelated sidebar change; the scrolling code and GPUI list revision are identical to the final candidate. All tests and timed measurements above include the sidebar fix.

Benchmarks used a dedicated Xvfb display, a 1280×800 foreground window, software Vulkan with four render workers, fresh local profiles, and composer submission. Compilation and recording were stopped during timed replays. This is a matched Linux comparison, not a native macOS/Metal benchmark or long-duration stability claim. Normal shared-host noise limits interpretation of small differences.

Reproduction

Build each revision with cargo build --release --locked -p zeron, copy each binary to an immutable path, and run scripts/resource-profile.mjs sequentially in main/candidate/candidate/main order for each workload. Use:

DISPLAY=:108 WAYLAND_DISPLAY= LP_NUM_THREADS=4 ZERON_FRAME_STATS=0 \
  ZERON_PROFILE_PSS=1 ZERON_PROFILE_SUBMIT_UI=1 \
  ZERON_PROFILE_PRE_IDLE_MS=10000 ZERON_PROFILE_IDLE_MS=45000 \
  ZERON_REPLAY_DELAY_MS=40 \
  CLAUDE_CODE_EXECUTABLE="$PWD/scripts/replay-claude.py" \
  ZERON_REPLAY_JOURNAL="$PWD/scripts/fixtures/resource-stream.jsonl" \
  node scripts/resource-profile.mjs /path/to/binary /tmp/fresh-run claude-code

The long fixture contains 52,624 text/reasoning bytes. For the short retained runway, use scripts/fixtures/runway-short-stream.jsonl, a 400 ms delta delay, 5 s pre-idle, and 20 s settled time. No model API calls are made.