Canvas snapshot & reconnection recovery

[FuJacob]

Context

The CRDT op log is currently in-memory only on the collab server (docOpLog in CanvasWebSocketHandler). Server restart = all CRDT history lost. Reconnecting clients get stale or empty state. Sim uses durable server-side snapshots + op logs for crash recovery and reconnection.

What to build

Server-side: Periodic snapshots

• Persist full canvas snapshots to PostgreSQL at regular intervals (e.g., every 30s of activity, or on significant structural changes)
• Snapshot = serialized state of all nodes + edges + CRDT document states
• Keep last N snapshots for rollback capability

Durable CRDT op log

• Move docOpLog from in-memory Map<String, List<JsonNode>> to PostgreSQL
• Move docSeenOpKeys dedup set to DB (or derive from op log)
• Index by (canvasId, docId, seq) for efficient range queries

Reconnection protocol

• Client sends sync_request with its stateVector (already implemented)
• Server responds with ops since that vector from durable storage (not in-memory)
• If gap is too large (client too far behind), send full snapshot instead

Crash recovery

• On server restart, load active canvas state from latest snapshot + pending ops from durable log
• Resume serving without data loss

Current code references

• services/collab/src/main/java/com/leetdoodle/collab/handler/CanvasWebSocketHandler.java — in-memory docOpLog, docSeenOpKeys, handleSyncRequest()
• frontend/src/shared/crdt/document.ts — getStateVector(), getOpsSince() (client-side sync protocol already exists)
• frontend/src/shared/crdt/useCanvasCrdt.ts — sync request hook

Acceptance criteria

• CRDT op log persisted to PostgreSQL
• Canvas snapshots saved periodically
• Reconnecting client receives missed ops from durable storage
• Server restart doesn't lose canvas state
• Large gaps trigger full snapshot sync instead of op replay