Web & Networking › Real-Time Communication

CRDT

Data structures that merge concurrent edits without conflicts, used in collaborative editors.

Also known as: crdt, conflict-free replicated data type, crdts

A CRDT (conflict-free replicated data type) is a data structure whose concurrent updates always merge to the same result — no coordination, no locking, no conflicts to resolve. Counters (grow-only, add wins), sets (add/remove with tombstones or observed semantics), registers (last-writer with logical clocks), and full collaborative-text types all have CRDT forms. Replicas exchange operations and converge automatically.

you: set title = "A"    peer: set title = "B"   (concurrent)
merge → same result everywhere (LWW, multi-value, or structured merge)

That’s the superpower for offline-first and realtime collaboration: edits apply locally instantly, sync in the background, and merge deterministically. The price is type design (not every structure has a good CRDT), metadata overhead (tombstones, clocks), and semantics that surprise (deleted items resurfacing, concurrent edits both surviving).

The classic mistakes:

• Using CRDTs where coordination is fine. A central server with operational transformation or simple locking is easier to reason about. CRDTs earn their complexity in peer-to-peer, offline-first, or partition-tolerant settings.
• Last-write-wins for everything. LWW registers silently drop concurrent edits — fine for presence, wrong for prose. Match the type to the collaboration semantics.
• Unbounded tombstones. Remove-wins sets accumulate tombstones forever without garbage collection; long-lived documents bloat. Plan compaction.
• Assuming text CRDTs are simple. Sequence CRDTs (RGA, Yjs-style) are subtle machinery — use proven libraries, never hand-roll.
• Ignoring identity and ordering. Concurrent inserts need stable, ordered identifiers; naive timestamps tie and interleave wrongly. The ID scheme is the algorithm.
• No server authority. Pure peer convergence can’t enforce permissions or validation. Real systems pair CRDT sync with server-side authz and bounds.
• Treating convergence as correctness. Replicas agreeing doesn’t mean the merged result is what users wanted. Test merged outcomes, not just convergence.

When to choose them: offline-first apps and peer collaboration where local-first writes matter more than central control. Otherwise, centralised approaches are simpler — see OT for the server-mediated classic.