Performance Audit

[PythonFZ]

Tier 1 — User-visible, multi-second cost per interaction

#	Finding	Evidence	Track
1	Initial bundle = 10.28 MB transferred on landing = 100 % of total gzipped bundle. Nothing is actually lazy-loaded despite manual chunks.	dockview audit: transfer ≈ sum of all chunk gzipped sizes (within 70 KB)	network/bundle
2	SmilesEditDialog (Ketcher) ships 6.7 MB gzip / 23 MB raw on every landing — only needed when editing molecules.	src/zndraw/static/assets/SmilesEditDialog-*.js	bundle
3	PUT /v1/rooms/{room_id}/step makes 212 sequential REST calls during 100-frame scrub = 1.4 s of round-trip wait per scrub (p50 14 ms × 100). Single biggest server-time consumer in the audit.	network-results.md §1, total 3.2 s across audit	network
4	TTI proxy = 3,489 ms in prod (canvas-visible). 2.65 s of that is post-FCP — bundle parse + hydration cost driven by 1.	2026-04-17-dockview-perf-audit-results.md headline	client perf
5	plotly ships 1.5 MB gzip on every landing despite being in its own chunk. modulePreload exclusion list isn't preventing the import.	bundle summary	bundle

Tier 2 — Hundreds of ms per interaction or per-session

#	Finding	Evidence	Track
6	room-switch triggers 117 React commits/iter — 3× cold-load baseline. 70 in dockview-layout, 47 in viewer. ~140 ms cumulative React work per switch.	react-results.md §1	react
7	GET /v1/joblib/rooms/{room_id}/providers polled 133× across 24 runs (~5.5/run). 2.1 s server time. No socket-driven invalidation.	network-results.md §2	network
8	GET /v1/rooms polled 100× (~4/run). 1.7 s total server time.	network-results.md	network
9	POST /v1/auth/guest p50 47 ms, p99 82 ms — blocks socket connect → blocks canvas mount on first visit. Function-level cause unknown (py-spy gap).	network-results.md §3	network
10	Box key={remountKey} in ViewerView.tsx:51 forces Canvas remount on every room change — likely root cause of 6. WebGL context destroy/recreate on top of React unmount cascade.	code inspection + react cascade signal	react
11	GET /v1/rooms/{room_id}/geometries returns 132 KB on every call — no ETag, no diff. 7.4 MB transferred per audit (56 calls × 132 KB).	network-results.md §4	network
12	POST /v1/rooms p99 = 133 ms, max 206 ms (vs p50 10 ms) — cold-path spike on first room create per server.	network-results.md §6	network
13	scipy imported eagerly at startup — scipy/integrate/_bvp.py 199 ms + scipy/special/__init__.py 180 ms = ~400 ms of cold-start time.	python-results.md "What flames did surface"	python
14	index.modern-BBJUX7oL.js (1.2 MB raw / 336 KB gzip) — second plotly fragment, also on landing.	bundle summary	bundle
15	index main chunk = 868 KB gzip — panels/registry.tsx imports every panel synchronously (GeometryPanel, RoomsPanel, PlotsBrowserPanel, FilesystemPanel, ChatPanel, SelectionsPanel, plus SecondaryPanel + ViewerView).	dockview audit §2	bundle

Tier 3 — Tens of ms per interaction, frequent enough to add up

#	Finding	Evidence	Track
16	mui chunk = 249 KB gzip fuses x-data-grid + x-tree-view with core MUI — grid + tree are panel-scoped (Geometries/Rooms only) but co-bundled with site-wide MUI.	dockview audit §3	bundle
17	open-rooms panel-open fires 50 ms long-task in prod / 346 ms in dev — DataGrid mount + /v1/rooms fetch on first open.	dockview audit "open-rooms"	client perf
18	open-geometries panel-open fires 51 ms long-task in prod / 239 ms in dev.	dockview audit	client perf
19	GET /v1/rooms/{room_id}/edit-lock polled 56× — 1.07 s total. Why is edit-lock fetched repeatedly per session?	network-results.md	network
20	GET /v1/rooms/{room_id}/frame-selection called 56× — 1.06 s total. Selection state currently REST-pulled, not socket-pushed.	network-results.md	network
21	PUT /.../sessions/{id}/active-camera called 43× — 927 ms total. Camera state syncs per interaction.	network-results.md	network
22	GET /v1/rooms/{room_id}/bookmarks called 56× — 873 ms. Refetched on most page loads.	network-results.md	network
23	GET /v1/rooms/{room_id}/selection-groups called 56× — 783 ms. Same pattern.	network-results.md	network
24	GET /v1/rooms/{room_id}/geometries/{key:path} called 56× — 690 ms.	network-results.md	network
25	socket.io + zndraw_socketio/wrapper.py add ~500 ms to startup.	python-results.md	python
26	include_router 257 ms during startup — 25 routers registered eagerly. Architectural.	python-results.md	python
27	<module> src/zndraw/app.py self-time 277 ms — top-level imports cascade.	python-results.md	python
28	resolve_or_refresh_token 176 ms during cold-load CLI — pre-flight token check is non-trivial.	python-results.md	python
29	selection flow triggers 51 React commits per iter for 5 clicks + 1 modifier (~9 commits per click).	react-results.md §4	react
30	geometry_invalidate socket event fires 3-15× per flow — high frequency, likely triggers React cascades.	network ws-frames	network
31	room_join fires 6-21× per flow — every page navigation triggers a fresh join, with full state re-fetch downstream.	network ws-frames	network

Tier 4 — Single-digit ms or one-shot startup costs

#	Finding	Evidence	Track
32	GET /v1/rooms/{room_id}/frames called 43× — 13 ms p50, 594 ms total.	network-results.md	network
33	GET /v1/rooms/{room_id}/frames/{index}/metadata called 43× — 7 ms p50, 349 ms.	network-results.md	network
34	GET /v1/server-settings/default-room called 36× — 5 ms p50, 305 ms. Frequently re-fetched.	network-results.md	network
35	GET /v1/auth/users/me called 70× — 3 ms p50, 283 ms. Re-checked constantly.	network-results.md	network
36	active_camera_update socket event fires 3-12× per flow — modest.	network ws-frames	network
37	Dev-mode heap = 139 MB after landing (2.6× prod) — HMR + source-map overhead. Dev only, not user-visible.	dockview audit	client perf
38	Dev-mode resource count = 250 (vs 33 prod) — per-module loads. Dev only.	dockview audit	client perf
39	Per React commit cost is consistently sub-ms (avg 0.5-0.9 ms). Good news — focus on count, not weight.	react-results.md §2	react
40	frame-scrub triggers only 39 React commits for 100 keypresses — R3F + zustand bypass works. Previously suspected "Fix F" (R3F subscription patterns from the dockview spec) is rejected on data.	react-results.md §3	react