CollabCanvas
Figma-like multiplayer canvas — real-time co-editing with LWW convergence, a Redis op journal, and honest documented trade-offs.
- Period
- Nov 2026
- Role
- System design & implementation
- Commits
- 150
- Key metric
- LWW + central seq · Redis stream journal · quadtree renderer
- Links
- Repo ↗
Problem
Real-time co-editing is the classic “looks easy, isn’t” problem. Textbooks say OT or CRDTs — Figma itself doesn’t use them. The study question: can a central ordering point plus per-property last-writer-wins give you correct convergence with a fraction of the machinery?
Architecture
One stateful room process owns each document — it IS the ordering point. Clients submit ops through a SignalR hub; the room takes a lock, dedups, assigns seq, applies (every op is idempotent and order-agnostic by construction), journals to a Redis Stream, then fans out on doc:out. Snapshots land every 100 ops — restart is snapshot + replay tail. Redis does four jobs: stream journal, snapshot/blob store, pub/sub transport, membership state. Canvas 2D with a quadtree keeps hit-testing and culling at O(visible), not O(objects).
Key decisions
- No OT, no full CRDT — a server that sees every write removes the merge problem; LWW registers converge by construction.
- Fractional index (base-62) for z-order — deterministic merge on every replica; float64 midpoints die after ~50 inserts, integers cause reindex storms.
- Optimistic UI with honest rebase — local ops apply instantly at a fractional temp seq; the server echo always wins; reconnect replays pending ops over fresh state.
- Owner-per-doc via Redis claim — submits to non-owners travel pub/sub to the owner. The split-brain window is documented, not hidden.
- Server sessions, not JWT — logout or revoke kills a live socket on the next op.
- Wire contract tested live — a Node 24 bot with the real
@microsoft/signalrclient verifies the protocol end to end; no jsdom theatre.
What it proves
Deliberate engineering under a deadline: docs/architecture.md writes the trade-offs down — the stack table lists every rejected alternative and why, and the split-brain window of the owner-claim model is documented, not hidden. The doc admits what the MVP doesn’t do — that’s the same honesty I apply to AI output.