Phase 6.2 vis-network Canvas Trial

Date: 2026-06-19 Branch: codex/large-graph-performance-experience Task: 6.2 Renderer: isolated vis-network Canvas trial

Scope

This report evaluates vis-network as the Canvas comparison candidate. It does not switch the production workbench graph path. Current DOM/SVG remains the rich small-graph and community-reading path until the Phase 6.4 route decision.

The trial uses the same graph shapes and interaction actions as the Sigma/Graphology trial. It also uses a shared graph-engine adapter contract so object ids, community ids, search hits, Pin hints, selected objects, and aggregation markers remain portable.

Commands

git log --oneline -15
npm run test --workspace=@llm-wiki/graph-engine
npm install -w @llm-wiki/graph-engine -D vis-network
node --import tsx --check tests/browser/graph-vis-network-trial.ts
node --import tsx --test packages/graph-engine/test/vis-network-trial-adapter.test.ts
GRAPH_VIS_TRIAL_ARTIFACT_DIR=/tmp/llm-wiki-graph-vis-trial-task-6-2 bash tests/graph-vis-network-trial.regression-1.sh

Artifacts

Machine-readable result:

/tmp/llm-wiki-graph-vis-trial-task-6-2/vis-network-trial-results.json

The artifact contains 47 fixed-schema records across 5 graph shapes.

Post-Review Harness Hardening

The original result table below is now treated as a historical isolation baseline. After review, the vis-network trial harness was hardened so a run fails when any action record fails, any failure class is present, any required action is missing, any requested shape is missing, or the wrapper only produced a JSON file without valid contents.

The default shape set was expanded from 5 shapes to the full 11-shape stress matrix: realistic proxy, 1000 sparse/dense, 5000 sparse/dense, 10000 aggregation/high-edge, oversized community, many small communities, many search hits, and many Pin nodes. Repeated interaction/memory cycles now run on every requested shape instead of only 1000-node shapes.

Interaction timing now waits for visible render completion after scripted actions. vis-network uses afterDrawing completion checks for render-affecting actions; the earlier timeout-based ready fallback was removed so a page cannot be marked ready without a completed draw. A hardened rerun is required before using these numbers as final production-integration evidence.

Result Table

Shape	Nodes	Edges	DOM nodes	Initial	Wheel	Pan	Search	Point	Container	Drawer	Community	Return	Cycle
nodes-1000-sparse	1000	1400	20	pass (117.6 ms)	pass (60.4 FPS)	pass (266.4 ms)	pass (28.0 ms)	pass (18.1 ms)	pass (9.3 ms)	pass (0.6 ms)	pass (7.3 ms)	pass (2.4 ms)	pass (0 MB growth)
nodes-1000-dense	1000	12000	20	pass (215.2 ms)	pass (60.4 FPS)	pass (292.0 ms)	pass (40.8 ms)	pass (49.4 ms)	pass (24.7 ms)	pass (0.6 ms)	pass (24.5 ms)	pass (8.4 ms)	pass (0 MB growth)
nodes-5000-sparse	5000	6500	20	pass (263.0 ms)	pass (60.2 FPS)	pass (274.3 ms)	pass (81.7 ms)	pass (66.3 ms)	pass (33.3 ms)	pass (0.4 ms)	pass (32.5 ms)	pass (9.8 ms)	not run
nodes-10000-aggregation	10000	14000	20	pass (456.5 ms)	pass (60.9 FPS)	pass (331.6 ms)	pass (148.0 ms)	pass (138.3 ms)	pass (91.0 ms)	pass (5.2 ms)	pass (94.6 ms)	pass (19.3 ms)	not run
oversized-community	3000	7000	20	pass (216.5 ms)	pass (60.5 FPS)	pass (270.5 ms)	pass (63.1 ms)	pass (46.4 ms)	pass (27.1 ms)	pass (0.5 ms)	pass (25.8 ms)	pass (7.0 ms)	not run

Comparison To Sigma And DOM/SVG

Shape	DOM/SVG initial	Sigma initial	vis-network initial	Sigma search	vis-network search	Sigma point	vis-network point
nodes-1000-sparse	900.4 ms	146.2 ms	117.6 ms	9.4 ms	28.0 ms	2.9 ms	18.1 ms
nodes-1000-dense	1008.4 ms	116.7 ms	215.2 ms	11.5 ms	40.8 ms	8.3 ms	49.4 ms
nodes-5000-sparse	4615.6 ms	175.3 ms	263.0 ms	23.4 ms	81.7 ms	13.0 ms	66.3 ms
nodes-10000-aggregation	8836.0 ms	289.1 ms	456.5 ms	40.9 ms	148.0 ms	14.6 ms	138.3 ms
oversized-community	15788.6 ms	140.6 ms	216.5 ms	14.7 ms	63.1 ms	6.9 ms	46.4 ms

Both candidate renderers remove the current DOM/SVG bottleneck class for global browsing. vis-network keeps wheel zoom near 60 FPS on all required shapes, but it is slower than Sigma/Graphology for search, point selection, and container/community updates in this isolated harness.

Integration Risks

vis-network owns more of the interaction and selection model by default. The trial disables physics and routes selection through explicit shared commands to avoid fighting llm-wiki drawer/community/search semantics.
Layout ownership can become opaque if production integration enables vis-network physics or stabilization. Pin and fixed-position behavior should remain owned by graph-engine, not by vis-network runtime state.
The Canvas route depends on DataSet mutation patterns. Search and selection are still acceptable in this trial, but they scale worse than the Sigma candidate on the same data.
Desktop packaging is plausible because the semantic layer remains outside the renderer, but production integration would need a strict adapter boundary to prevent product logic from moving into vis-network callbacks.

Interpretation

vis-network is a viable Canvas comparison candidate, but this trial does not currently beat Sigma/Graphology on the important semantic-update actions. It remains in the Phase 6 comparison set until the aggregation fallback and final route decision are recorded.

Acceptance Evidence

Historical shapes measured: 1000 sparse, 1000 dense, 5000 sparse, 10000 aggregation, and oversized-community.
Hardened default shapes now include the full 11-shape stress matrix.
Required actions are now enforced for every requested shape: initial render, pan, zoom, search highlight, point select, container select, drawer open, enter community, return global, and repeated memory cycle.
Behavior parity test passed: node --import tsx --test packages/graph-engine/test/vis-network-trial-adapter.test.ts.
No production workbench renderer path was switched.

6.2 KiB Raw Permalink Blame History