Files
pipelines/app/notes/perf-round3.md
T
alan f1ac647125 perf: build unshown run-detail tabs after the route transition; interaction bench
FadeThroughStack(lazy: true) builds only the shown page first and the rest
one per frame once no animation runs. Opening a run detail no longer lays out
Timeline, Artifacts and Changes offstage during the route transition
(transition build 62-103 ms -> 38-59 ms); first tab switches are unchanged.

Adds tool/perf/interaction_bench.mjs (FrameTiming per interaction over the
full-app fixture, CPU profiles mapped to Dart files), fixture options for
hits/chats/graph, and notes/perf-round3.md with the survey.
2026-11-02 13:54:56 +00:00

6.3 KiB

Performance round 3: survey and lazy run-detail tabs

This round started from measurements. It follows notes/chat-stream-perf.md, notes/chat-open-perf.md and notes/perf-glass-chrome.md.

Tools

  • tool/perf/interaction_bench.mjs runs scripted interactions over the release full-app fixture (tool/search_selection_probe_entry.dart) at 560x900, Material, light theme. It clicks points, so Flutter semantics stay off. It records Flutter FrameTiming (tool/perf/frame_log.dart, ?perf=1) and prints medians: build sum and max, raster sum and max, and the longest frame. --frames prints every frame. --cpuprofile together with a --profile --source-maps build assigns V8 samples to Dart files (tool/perf/cpu_attribution.mjs).
  • tool/perf/perf_fixture.dart adds fixture data: ?hits=N Search hits, ?chats=N chats (each opens a short markdown transcript), and ?graph=1 (run 7 includes its graph).
  • tool/perf/js_size_by_source.mjs splits main.dart.js by Dart source.

Headless Chrome has no GPU, so CanvasKit rasterises on the CPU and raster time dominates every frame (see the glass note: with the GPU, frames take about 2-3 ms). Build time is the part that the app's code controls.

Survey (base main ec619234, medians of 5)

Ranked by the worst build frame (ms):

interaction build max build sum longest frame
Chat tab, first visit, 150 chats 116.5 189 143
Settings, from Pipelines 33.6 65 46
Run detail open, with graph preview 27-30 83-91 60-67
Run detail open (runs list, 2 or 400 runs) 24-29 72-90 55-66
Search, typing release (209 hits) 27.5 127 (24 frames) 48
Chat tab, first visit, 30 chats 27-35 74-83 52-58
Pipeline detail open 16-22 55-66 56-60
Run graph full screen 20.8 35 65
Chat list to open chat (20 messages) 20.3 75 61
Run Timeline tab, 400 steps 17-25 36-46 33-44
Main tab switch (Runs/Pipelines/Search) 15-19 50-60 40-46
Visit page from Timeline 12-20 36-49 54-59
Run page tabs (Summary/Artifacts/Changes) 10-18 28-37 35-40
Pipeline Graph and Source tabs 8-16 27-37 37-42
Artifact viewer open (the fixture serves no file) 4.3 10 29
Startup 9.4 37 (15 frames) 27

Startup: the first rasterised frame comes 300-400 ms after navigation with a warm cache and about 1.5 s cold. A cold load spends about 1.15 s fetching and compiling canvaskit.wasm. main.dart.js is 4.0 MiB (1.2 MiB gzipped), and app code is about 15% of it (chat 171 KiB, the largest feature). Deferred loading would save well under 100 KiB gzipped per feature, so it is not worth doing. The icon fonts are already tree-shaken (MaterialIcons 1.6 MB to 18 KB). Startup does no heavy synchronous work.

Fix: lazy run-detail tabs

Profiling showed that opening a run detail built and laid out every page tab offstage (Timeline, Artifacts, Changes) during the route transition, because FadeThroughStack mounts all of its children. FadeThroughStack(lazy: true) now builds only the shown page at first. It builds the rest one per frame once no animation is running (no transient frame callbacks, so after the route transition), or at once when one is shown first. After that, every page stays mounted, as before. Only RunDetailTabs sets lazy. Making pipeline detail and visit-page sections lazy gave no difference beyond noise, so those stay as they were.

Before and after, runs->run detail, three runs each with --frames. Each cell is the build time of the frames during the route transition (the first 9 frames):

before after
transition build sum 62-103 ms 38-59 ms
worst transition build frame 20.7-54.1 ms 16-34 ms

Medians of 5, two rounds alternating with main:

before after
open: build max 25.6 / 26.4 22.9 / 19.2 (now an idle prebuild frame)
open, 400 runs: build max 28.9 / 28.5 21.0 / 22.5
open with graph preview: build max 25.3 / 28.2 21.2
first switch to Timeline: build max 13.0 / 9.8 14.1 / 16.3 (noise)
first switch to Timeline, 400 steps 17.5 / 17.3 18.0
Artifacts, Changes switches 14.6-16.1 9.4-15.1

The build sum for the open window goes up by about 10 ms, because the pages built later now take frames of their own. Those frames come after the transition, while nothing animates. A first attempt that built pages only when shown made the open cheaper (build max about 16 ms), but it moved the cost to the first switch to Timeline (build max 17 to 24 ms, 400 steps 21 to 28 ms). The idle prebuild avoids that.

Behaviour: no visible change. Within the first frames after a run detail opens, a tab the operator has not shown is not mounted yet. Its offstage widgets (for example the artifact gallery's thumbnails) therefore start a few frames later. If an animation runs without stopping, unshown tabs are built when they are first shown.

Candidates for round 4

  1. Chat list rows (largest value). MaterialSessionList and CupertinoSessionList build every row of an open section in a Column inside CollapsibleListSection's AnimatedSize. The first Chat visit costs about 0.75 ms of build per chat: 35 ms at 30 chats and 116 ms at 150, which is a dropped-frame burst during the tab fade. Moving to slivers (heading plus a lazy SliverList) fixes the scaling, but the section's size animation has to be rebuilt, both for open and close (explicit Align.heightFactor) and for rows arriving into an open section. Without that it is a visual change, so it needs its own careful item.
  2. Settings (33 ms build frame). Everything sits in one SingleChildScrollView (about 1350 px of content). A lazy list would dispose offscreen sections that own controllers, so it needs keep-alives or lifted state. Medium value: the page is opened rarely.
  3. Run detail open, other costs (about 15-20 ms left): the header plus Summary, and detailTimeline and other step derivations that _body still runs for unbuilt tabs on every rebuild. Builders instead of child widgets would remove that work. Low to medium value.
  4. Search result frame (27 ms for the first page of rows). Low value.
  5. Cold start: canvaskit.wasm makes up most of the 1.5 s cold first frame. Only caching and service-worker policy could help. Out of the app's control.